2512.20237

# MemR3: Memory Retrieval via Reflective Reasoning for LLM Agents **Authors**: Xingbo Du, Loka Li, Duzhen Zhang, Le Song Abstract Memory systems have been designed to leverage past experiences in Large Language Model (LLM) agents. However, many deployed memory systems primarily optimize compression and storage, with comparatively less emphasis on explicit, closed-loop control of memory retrieval. From this observation, we build memory retrieval as an autonomous, accurate, and compatible agent system, named MemR 3, which has two core mechanisms: 1) a router that selects among retrieve, reflect, and answer actions to optimize answer quality; 2) a global evidence-gap tracker that explicitly renders the answering process transparent and tracks the evidence collection process. This design departs from the standard retrieve-then-answer pipeline by introducing a closed-loop control mechanism that enables autonomous decision-making. Empirical results on the LoCoMo benchmark demonstrate that MemR 3 surpasses strong baselines on LLM-as-a-Judge score, and particularly, it improves existing retrievers across four categories with an overall improvement on RAG (+7.29%) and Zep (+1.94%) using GPT-4.1-mini backend, offering a plug-and-play controller for existing memory stores. Machine Learning, ICML 1 Introduction With recent advances in large language model (LLM) agents, memory systems have become the focus of storing and retrieving long-term, personalized memories. They can typically be categorized into two groups: 1) Parametric methods (wang2024wise; fang2025alphaedit) that encode memories implicitly into model parameters, which can handle specific knowledge better but struggle in scalability and continual updates, as modifying parameters to incorporate new memories often risks catastrophic forgetting and requires expensive fine-tuning. 2) Non-parametric methods (xu2025amem; langmem_blog2025; chhikara2025mem0; rasmussen2025zep), in contrast, store explicit external information, enabling flexible retrieval and continual augmentation without altering model parameters. However, they typically rely on heuristic retrieval strategies, which can lead to noisy recall, heavy retrieval, and increasing latency as the memory store grows. Orthogonal to these works, this paper constructs an agentic memory system, MemR 3, i.e., Mem ory R etrieval system with R eflective R easoning, to improve retrieval quality and efficiency. Specifically, this system is constructed using LangGraph (langchain2025langgraph), with a router node selecting three optional nodes: 1) the retrieve node, which is based on existing memory systems, can retrieve multiple times with updated retrieval queries. 2) the reflect node, iteratively reasoning based on the current acquired evidence and the gaps between questions and evidence. 3) the answer node that produces the final response using the acquired information. Within all nodes, the system maintains a global evidence-gap tracker to update the acquired (evidence) and missing (gap) information. <details> <summary>x1.png Details</summary>  ### Visual Description ## Diagram Type: Comparative Workflow for Memory-Augmented LLMs ### Overview This technical diagram compares three different approaches for a Large Language Model (LLM) to answer a specific temporal reasoning question based on a set of "memories" (stored conversational data). The diagram is organized into four main panels: a top-left query box, two left-side panels showing failed baseline methods, and two large panels on the right detailing the successful iterative process of a method called **MemR³**. ### Components/Axes * **Query $q$ (Top-Left):** Defines the input task. * **1) Full-Context (Middle-Left):** A baseline method that provides all available data to the model at once. * **2) Retrieve-then-Answer (Bottom-Left):** A standard Retrieval-Augmented Generation (RAG) approach. * **3) MemR³ (Center and Right):** A multi-step, iterative reasoning and retrieval framework, split into two parts (1/2 and 2/2). * **Color Coding:** * **Blue Text in Brackets:** Represents specific system actions or "Acts" (e.g., `Retrieve`, `Reflect`, `Answer`). * **Orange Highlight:** Indicates incorrect answers or failed outputs. * **Green Highlight:** Indicates successful evidence extraction, gap identification, query refinement, and the final correct answer. * **Red 'X' Icon:** Denotes a failure in performance or accuracy. * **Green Checkmark Icon:** Denotes a successful and correct output. --- ### Content Details #### Header: Query $q$ * **Question:** "How many months passed between Andrew adopting Toby and Buddy?" #### Panel 1: Full-Context * **Action:** `Retrieve[all memories]` * **Answer (Orange):** "Six months passed between Andrew adopting Toby and Buddy. [wrong]" * **Annotation (Red):** "Heavy Context reduces LLM's performance" (accompanied by a red 'X'). #### Panel 2: Retrieve-then-Answer * **Action:** `Retrieve[q-relevant memories]` * **Memories Extracted:** 1. `[11 July, 2023] Andrew: Hey! So much has changed since last time we talked - meet Toby, my puppy.` 2. `[19 October, 2023] Andrew: Speaking of which, I've got some awesome news - I recently adopted another pup from a shelter. He's the best.` 3. `...` * **Answer (Orange):** "About 4 months. [wrong]" (accompanied by a red 'X'). #### Panel 3: MemR³ (1/2) * **Act 0:** `Retrieve[q-relevant memories]` * **Memories:** `[the same as that in 2)]` * **Evidence (Green):** "Andrew adopted Toby on July 11, 2023, and another pup was adopted near October 19, 2023." * **Gaps (Green):** "The name of another pup is unknown, and it lacks the specific adoption date for Buddy." * **Action:** `Retrieve` * **Refined Query $\Delta q$ (Green):** "Buddy adoption date" * **New Query Formula:** $q^{ret} = q \oplus \Delta q$ * **Act 1:** `Retrieve[q^{ret}-relevant memories]` * **Memories Extracted:** 1. `[19 October, 2023] Andrew: I named him Buddy because he's my buddy and I hope him and Toby become buddies!` 2. `...` * **Updated Evidence (Green):** "Andrew adopted Toby on July 11, 2023, and Buddy was named on October 19, 2023." * **Updated Gaps (Green):** "It lacks the specific adoption date for Buddy." #### Panel 4: MemR³ (2/2) * **Action:** `Reflect` * **Reasoning:** "Though it lacks the specific adoption date for Buddy, we can calculate the approximate number of months between the two events." * **Act 2:** `Reflect` * **Updated Evidence/Gaps:** `[The same as above].` * **Action:** `Answer` * **Draft Answer:** "To determine how many months passed between Andrew adopting Toby and Buddy, we need to look at the specific dates of these events. 1. Andrew adopted Toby on **July 11, 2023** 2. Buddy was named on **October 19, 2023**. Now, we calculate the time between these two dates: *[Calculation process omitted]* Therefore, the total number of full months that passed between Andrew adopting Toby and Buddy is **3 months**." * **Act 3:** `Answer` * **Final Answer (Green):** "3 months. [correct]" (accompanied by a green checkmark). --- ### Key Observations * **Context Overload:** The "Full-Context" method fails significantly (stating 6 months), suggesting that providing too much irrelevant data confuses the model's reasoning capabilities. * **Precision Errors:** The "Retrieve-then-Answer" method gets closer (4 months) but fails on the exact calculation, likely due to a lack of explicit reasoning steps. * **Iterative Refinement:** MemR³ identifies that the initial retrieval didn't explicitly link the name "Buddy" to the second adoption event. It generates a refined query ($\Delta q$) to confirm the identity of the second dog. * **Explicit Reasoning:** MemR³ includes a "Reflect" stage where it acknowledges missing data (exact adoption date for Buddy) but determines that the naming date is a sufficient proxy for calculation. ### Interpretation The data demonstrates that for complex temporal reasoning tasks, a **multi-hop, iterative retrieval and reflection process** (MemR³) is superior to single-pass methods. 1. **The "Heavy Context" Problem:** The diagram suggests that LLMs have a "distraction" threshold. When forced to process all memories, the model's accuracy drops more than when it uses a targeted retrieval. 2. **Gap Analysis as a Catalyst:** The critical innovation in MemR³ shown here is the explicit identification of "Gaps." By articulating what it *doesn't* know, the system can perform a second, targeted retrieval to bridge the information void. 3. **Logical Proxying:** In the final stage, the model uses the naming date (Oct 19) as a proxy for the adoption date. The calculation (July 11 to Oct 19) results in 3 full months (July-Aug, Aug-Sept, Sept-Oct). The "Retrieve-then-Answer" method likely rounded up to 4 months, whereas MemR³'s explicit "Draft Answer" process ensured a more precise calculation of "full months." </details> Figure 1: Illustration of three memory-usage paradigms. Full-Context overloads the LLM with all memories and answers incorrectly; Retrieve-then-Answer retrieves relevant snippets but still miscalculates. In contrast, MemR 3 iteratively retrieves and reflects using an evidence–gap tracker (Acts 0–3), refines the query about Buddy’s adoption date, and produces the correct answer (3 months). The system has three core advantages: 1) Accuracy and efficiency. By tracking the evidence and gap, and dynamically routing between retrieval and reflection, MemR 3 minimizes unnecessary lookups and reduces noise, resulting in faster, more accurate answers. 2) Plug-and-play usage. As a controller independent of existing retriever or memory storage, MemR 3 can be easily integrated into memory systems, improving retrieval quality without architectural changes. 3) Transparency and explainability. Since MemR 3 maintains an explicit evidence-gap state over the course of an interaction, it can expose which memories support a given answer and which pieces of information were still missing at each step, providing a human-readable trace of the agent’s decision process. We compare MemR 3, the Full-Context setting (which uses all available memories), and the commonly adopted retrieve-then-answer paradigm from a high-level perspective in Fig. 1. The contributions of this work are threefold in the following: (1) A specialized closed-loop retrieval controller for long-term conversational memory. We propose MemR 3, an autonomous controller that wraps existing memory stores and turns standard retrieve-then-answer pipelines into a closed-loop process with explicit actions (retrieve / reflect / answer) and simple early-stopping rules. This instantiates the general LLM-as-controller idea specifically for non-parametric, long-horizon conversational memory. (2) Evidence–gap state abstraction for explainable retrieval. MemR 3 maintains a global evidence–gap state $(\mathcal{E},\mathcal{G})$ that summarizes what has been reliably established in memory and what information remains missing. This state drives query refinement and stopping, and can be surfaced as a human-readable trace of the agent’s progress. We further formalize this abstraction via an abstract requirement space and prove basic monotonicity and completeness properties, which we later use to interpret empirical behaviors. (3) Empirical study across memory systems. We integrate MemR 3 with both chunk-based RAG and a graph-based backend (Zep) on the LoCoMo benchmark and compare it with recent memory systems and agentic retrievers. Across backends and question types, MemR 3 consistently improves LLM-as-a-Judge scores over its underlying retrievers. 2 Related Work 2.1 Memory for LLM Agents Prior work on non-parametric agent memory systems spans a wide range of fields, including management and utilization (du2025rethinking), by storing structured (rasmussen2025zep) or unstructured (zhong2024memorybank) external knowledge. Specifically, production-oriented agents such as MemGPT (packer2023memgpt) introduce an OS-style hierarchical memory system that allows the model to page information between context and external storage, and SCM (wang2023enhancing) provides a controller-based memory stream that retrieves and summarizes past information only when necessary. Additionally, Zep (rasmussen2025zep) builds a temporal knowledge graph that unifies and retrieves evolving conversational and business data. A-Mem (xu2025amem) creates self-organizing, Zettelkasten-style memory that links and evolves over time. Mem0 (chhikara2025mem0) extracts and manages persistent conversational facts with optional graph-structured memory. MIRIX (wang2025mirix) offers a multimodal, multi-agent memory system with six specialized memory types. LightMem (fang2025lightmem) proposes a lightweight and efficient memory system inspired by the Atkinson–Shiffrin model. Another related approach, Reflexion (shinn2023reflexion), improves language agents by providing verbal reinforcement across episodes by storing natural-language reflections to guide future trials. In this paper, we explicitly limit our scope to long-term conversational memory. Existing parametric approaches (wang2024wise; fang2025alphaedit), KV-cache–based mechanisms (zhong2024memorybank; eyuboglu2025cartridges), and streaming multi-task memory benchmarks (wei2025evo) are out of scope for this work. Orthogonal to existing storage, MemR 3 is an autonomous retrieval controller that uses a global evidence–gap tracker to route different actions, enabling closed-loop retrieval. 2.2 Agentic Retrieval-Augmented Generation Retrieval-Augmented Generation (RAG) (lewis2020retrieval) established the modern retrieve-then-answer paradigm; subsequent work explored stronger retrievers (karpukhin2020dense; izacard2021leveraging). Beyond the RAG, recent work, such as Self-RAG (asai2024self), Reflexion (shinn2023reflexion), ReAct (yao2022react), and FAIR-RAG (asl2025fair), has shown that letting a language model (LM) decide when to retrieve, when to reflect, and when to answer can substantially improve multi-step reasoning and factuality in tool-augmented settings. MemR 3 follows this general “LLM-as-controller” paradigm but applies it specifically to long-term conversational memory over non-parametric stores. Concretely, we adopt the idea of multi-step retrieval and self-reflection from these frameworks, but i) move the controller outside the base LM as a LangGraph program, ii) maintain an explicit evidence–gap state that separates verified memories from remaining uncertainties, and iii) interface this state with different memory backends (e.g., RAG and Zep (rasmussen2025zep)) commonly used in long-horizon dialogue agents. Our goal is not to replace these frameworks, but to provide a specialized retrieval controller that can be plugged into existing memory systems. <details> <summary>x2.png Details</summary>  ### Visual Description ## Diagram Type: LLM-based Retrieval and Reasoning Pipeline ### Overview This technical diagram illustrates an iterative workflow for a Large Language Model (LLM) system designed to answer complex queries by retrieving information from memory, identifying knowledge gaps, and reasoning through steps. The process is governed by a "Router" that manages the iteration budget and decides whether to continue retrieving information or finalize an answer. ### Components/Axes The diagram is organized into four primary functional regions from left to right: 1. **Input Data (Top-Left):** * **User Query $q$:** A text box containing the example question: *"How many months passed between Andrew adopting Toby and Buddy?"* * **Memory $\mathcal{M}$:** A container showing two types of data storage: * **Chunk-based:** Represented by a database icon and a stack of five horizontal bars (progress-style indicators) of varying lengths. * **Graph-based:** Represented by a network graph icon with nodes and connecting edges. 2. **LLM Generation (Center):** * **Evidence-gap Tracker:** A grey-headed box that splits information into: * **Evidence:** *"Andrew adopted Toby on July 11, 2023, and Buddy was named on October 19, 2023."* * **Gap:** *"It lacks the specific adoption date for Buddy."* * **Generated Actions:** A white box containing three potential outputs: * **Retrieve:** new query $\Delta q$ (Magnifying glass icon) * **Reflect:** reasoning $r$ (Thought bubble icon) * **Answer:** draft answer $w$ (Pen icon) 3. **Router (Bottom-Right):** * A control unit that manages the flow based on three internal checks: * **Iteration budget** (Four-way arrow icon) * **Reflect-streak capacity** (Battery/recharge icon) * **Retrieval opportunity check** (Checkbox icon) * **Control Buttons:** Three dark green/blue buttons sit above the router: **Reflect**, **Answer**, and **End**. 4. **Final Answer (Top-Right):** * A light green box containing the terminal output: *"Answer: the total number of full months that passed between Andrew adopting Toby and Buddy is 3 months."* --- ### Content Details #### Process Flow and Logic * **Initialization:** The process begins at the **Start** button (dark blue) which triggers the **Retrieve** action (dark green). * **Retrieval Phase:** The system pulls from **Memory $\mathcal{M}$** (both chunk and graph-based). The input to the LLM is denoted as $(q^{ret}, \mathcal{M})$. * **Generation Phase:** The LLM analyzes the retrieved data. In the provided example, it identifies that it knows Toby's adoption date but only Buddy's naming date, marking the adoption date for Buddy as a "Gap." * **Action Selection:** The LLM generates potential actions: a refined query ($\Delta q$), reasoning ($r$), or a draft answer ($w$). * **Routing & Feedback Loop:** * The **Router** evaluates the generated actions against its constraints (budget, streak, opportunities). * **Feedback Loop (Dashed Line):** If more information is needed, a dashed line carries a refined query $q^{ret} = q \oplus \Delta q$ back to the **Retrieve** step. * **Finalization:** If the reasoning is sufficient, the flow moves through the **Reflect** ($r$) or **Answer** ($w$) paths to the **Final Answer** box. * **Termination:** Once the **Final Answer** is produced, the process moves to the **End** state. --- ### Key Observations * **Iterative Refinement:** The formula $q^{ret} = q \oplus \Delta q$ indicates that the system doesn't just search once; it appends or modifies the original query based on discovered gaps to perform targeted follow-up retrievals. * **Hybrid Memory:** The system utilizes both unstructured (Chunk-based) and structured (Graph-based) memory, suggesting a RAG (Retrieval-Augmented Generation) architecture that can handle both semantic search and relational data. * **Guardrails:** The "Router" contains specific logic to prevent infinite loops ("Reflect-streak capacity") and manage computational costs ("Iteration budget"). --- ### Interpretation The data demonstrates a **Self-Correction/Reasoning Loop** in an AI agent. Rather than providing a hallucinated or incomplete answer when faced with missing data (the "Gap" regarding Buddy's adoption date), the system is designed to recognize what it *doesn't* know. The "Evidence-gap Tracker" acts as a logical bridge; it forces the model to explicitly state its premises before concluding. The final answer of "3 months" (calculating from July 11 to October 19) suggests that through the iterative retrieval loop ($\Delta q$), the system likely found the missing adoption date or determined that the naming date was the relevant proxy, allowing it to complete the reasoning chain. This architecture prioritizes accuracy and transparency over simple one-shot generation. </details> Figure 2: Pipeline of MemR 3. MemR 3 transforms retrieval into a closed-loop process: a router dynamically switches between Retrieve, Reflect, and Answer nodes while a global evidence–gap tracker maintains what is known and what is still missing. This enables iterative query refinement, targeted retrieval, and early stopping, making MemR 3 an autonomous, backend-agnostic retrieval controller. 3 MemR 3 In this section, we first formulate the problem and provide preliminaries in Sec. 3.1, and then give a system overview of MemR 3 in Sec. 3.2. Additionally, we describe the two core components that enable accurate and efficient retrieval: the router and the global evidence-gap tracker in Sec. 3.4 and Sec. 3.3, respectively. 3.1 Problem Formulation and Preliminaries We consider a long-horizon LLM agent that interacts with a user, forming a memory store $\mathcal{M}=\{m_{i}\}_{i=1}^{N}$ , where each memory item $m_{i}$ may correspond to a dialogue utterance, personal fact, structured record, or event, often accompanied by metadata such as timestamps or speakers. Given a user query $q$ , a retriever is applied to retrieve a set of memory snippets $\mathcal{S}$ that are useful for generating the final answer. Then, given designed prompt template $p$ , the goal is to produce an answer $w$ : $$ \begin{split}\mathcal{S}&\leftarrow\texttt{Retrieve}(q,\mathcal{M}).\\ w&\leftarrow\texttt{LLM}(q,\mathcal{S},p),\end{split} \tag{1} $$ which is accurate (consistent with all relevant memories in $\mathcal{M}$ ), efficient (requiring minimal retrieval cycles and low latency), and robust (stable under noisy, redundant, or incomplete memory stores) as much as possible. Existing memory systems have done great work on the memory storage $\mathcal{M}$ , but typically follow an open-loop pipeline: 1) apply a single retrieval pass; 2) feed the selected memories $\mathcal{S}$ into a generator to produce $\mathcal{A}$ . This approach lacks adaptivity: retrieval does not incorporate intermediate reasoning, and the system never represents which information remains missing. This leads to both under-retrieval (insufficient evidence) and over-retrieval (long, noisy contexts). MemR 3 addresses these limitations by treating retrieval as an autonomous sequential decision process with explicit modeling of both acquired evidence and remaining gaps. 3.2 System Overview MemR 3 is implemented as a directed agent graph comprising three operational nodes (Retrieve, Reflect, Answer) and one control node (Router) using LangGraph (langchain2025langgraph) (an open-source framework for building stateful, multi-agent workflows as graphs of interacting nodes). The agent maintains a mutable internal state $$ s=(q,\mathcal{S},\mathcal{E},\mathcal{G},k), \tag{2} $$ where $q$ and $\mathcal{S}$ are the aforementioned original user query and retrieved snippets, respectively. $\mathcal{E}$ is the accumulated evidence relevant to $q$ and $\mathcal{G}$ is the remaining missing information (the “gap”) between $q$ and $\mathcal{E}$ . Moreover, we maintain the iteration index $k$ to control early stopping. At each iteration $k$ , the router chooses an action in $\{\texttt{retrieve},\ \texttt{reflect},\ \texttt{answer}\}$ , which determines the next node in the computation graph. The pipeline is shown in Fig. 2. This transforms the classical retrieve-then-answer pipeline into a closed-loop controller that can repeatedly refine retrieval queries, integrate new evidence, and stop early once the information gap is resolved. 3.3 Global Evidence-Gap Tracker A core design principle of MemR 3 is to explicitly maintain and update two state variables: the evidence $\mathcal{E}$ and the gap $\mathcal{G}$ . These variables summarize what the agent currently knows and what it still needs to know to answer the question. At iteration $k$ , the evidence $\mathcal{E}_{k}$ and gaps $\mathcal{G}_{k}$ are updated according to the retrieved snippets $\mathcal{S}_{k-1}$ (from the retrieve node) or reflective reasoning $\mathcal{F}_{k-1}$ (from the reflect node), together with last evidence $\mathcal{E}_{k-1}$ and gaps $\mathcal{G}_{k-1}$ at $k-1$ iteration: $$ \mathcal{E}_{k},\mathcal{G}_{k},a_{k}=\texttt{LLM}(q,\mathcal{S}_{k-1},\mathcal{F}_{k-1},\mathcal{E}_{k-1},\mathcal{G}_{k-1},p_{k}), \tag{3} $$ where $p_{k}$ is the prompt template at $k$ iteration. Additionally, $a_{k}$ is the action at $k$ iteration, which will be introduced in Sec. 3.4. Note that we explicitly clarify in $p_{k}$ that $\mathcal{E}_{k}$ does not contain any information in $\mathcal{G}_{k}$ , making evidence and gaps decoupled. An example is shown in Fig. 3 to illustrate the evidence-gap tracker. <details> <summary>x3.png Details</summary>  ### Visual Description ## Diagram Type: Query-Response Logic Flow ### Overview This image displays a user interface (UI) component representing a structured interaction between a user and an information retrieval system (likely an AI assistant). It illustrates a "reasoning" process where a natural language query is analyzed, and the response is broken down into verified evidence and identified missing information. ### Components/Axes The image is divided into two primary horizontal blocks, one for the input and one for the output. * **Header Labels:** Both main blocks feature a centered, grey rectangular label at the top of their respective borders. * **Top Label:** "Query" * **Bottom Label:** "Response" * **Icons:** * **Query Icon (Top-Left):** A green icon depicting a person's silhouette with a speech bubble containing a question mark. * **Evidence Icon (Middle-Left):** A blue circular icon containing a magnifying glass with a pulse/waveform line inside. * **Gaps Icon (Bottom-Left):** A red icon consisting of three interlocking circles, similar to a Venn diagram. * **Text Blocks:** * **Query Text:** Located in the top box. * **Evidence Text:** Located in the upper half of the bottom box. * **Gaps Text:** Located in the lower half of the bottom box. ### Content Details The text within the diagram is transcribed exactly as follows: * **Query Section:** * "What happened 2 days after my last dentist appointment?" * **Response Section:** * **Evidence:** "You had a dentist appointment on July 12." * **Gaps:** "Information about events on July 14 is missing. Whether July 12 is indeed the most recent dentist appointment is unknown." ### Key Observations * **Temporal Reasoning:** The system demonstrates the ability to perform date arithmetic. It identifies a "dentist appointment" on July 12 and correctly calculates that "2 days after" corresponds to July 14. * **Uncertainty Handling:** The system explicitly identifies two types of missing information: 1. **Data Void:** No records exist for the calculated target date (July 14). 2. **Contextual Ambiguity:** The system cannot verify if the July 12 appointment is the *absolute* last one in the user's history, acknowledging a potential limitation in its data access or the user's records. * **Color Coding:** The use of blue for "Evidence" suggests a neutral or positive confirmation of facts, while red for "Gaps" serves as a warning or indicator of missing critical data. ### Interpretation This diagram demonstrates a "Transparent Reasoning" or "Grounded" AI response model. Instead of providing a simple "I don't know" or potentially hallucinating an answer, the system exposes its internal logic to the user. 1. **Fact Retrieval:** It successfully finds a relevant anchor point (July 12). 2. **Logic Application:** It applies the user's requested offset (+2 days). 3. **Validation:** It checks its database for the resulting date (July 14) and finds nothing. 4. **Critical Thinking:** It questions its own premise—is July 12 actually the "last" appointment? This approach is characteristic of systems designed for high-stakes personal data management (like health or scheduling), where accuracy and the ability to audit the system's "thought process" are more important than a concise but potentially misleading answer. It follows a Peircean investigative style by presenting the "Firstness" of the raw query, the "Secondness" of the conflicting data/gaps, and the "Thirdness" of the structured explanation that mediates between the two. </details> Figure 3: Example of the evidence-gap tracker for a specific query. At each step, the agent maintains an explicit summary of the evidence established and the information still missing. This state can be presented directly to users as a human-readable explanation of the agent’s progress in answering the query. Through the evidence-gap tracker, MemR 3 maintains a structured and transparent internal state that continuously refines the agent’s understanding of both i) what has already been established as relevant evidence, and ii) what missing information still prevents a complete and faithful answer. This explicit decoupling enables MemR 3 to reason under partial observability: as long as $\mathcal{G}_{k}≠\varnothing$ , the agent recognizes that its current knowledge is insufficient and can proactively issue a refined retrieval query to close the remaining gap. Conversely, when $\mathcal{G}_{k}$ becomes empty, the router detects that the agent has accumulated adequate evidence and can safely transition to the answer node. Beyond guiding retrieval, the evidence-gap representation also makes the agent’s behavior more transparent. At any iteration $k$ , the pair $(\mathcal{E}_{k},\mathcal{G}_{k})$ can be surfaced as a structured explanation of i) which memories the agent currently treats as relevant evidence and ii) which unresolved questions or missing details are preventing a confident answer. This trace provides users and developers with a faithful view of how the agent arrived at its final answer and why additional retrieval steps were taken (or not). In the following, we display an informal theorem that indicates the properties of the idealized evidence-gap tracker. **Theorem 3.1 ([Informal]Monotonicity, soundness, and completeness of the idealized evidence-gap tracker)** *Under an idealized requirement space $R(q)$ for a specific query $q$ , the evidence-gap tracker in MemR 3 is monotone (evidence never decreases and gaps never increase), sound (every supported requirement eventually enters the evidence set), and complete (if every requirement $r∈ R(q)$ is supported by some memory, the ideal gap eventually becomes empty).* Formally, in Appendix B we define the abstract requirement space $R(q)$ and characterize the tracker as a set-valued update on $R(q)$ , proving fundamental soundness, monotonicity, and completeness properties (Theorem B.4), which we later use in Sec. 4.3 to interpret empirical phenomena such as why some questions cannot be fully resolved even after exhausting the iteration budget. 3.4 LangGraph Nodes We explicitly define several nodes in the LangGraph framework, including start, end, generate, router, retrieve, reflect, answer. Specifically, start is always followed by retrieve, and end is reached after answer. generate is a LLM generation node, which is already introduced in Eq. 3. In the following, we further introduce the router node and three action nodes. Router. At each iteration, the router, an autonomous sequential controller, uses the current state and selects an action from $\{\texttt{retrieve},\texttt{reflect},\texttt{answer}\}$ . Each action $a_{k}$ is accompanied by a textual generation: $$ { a_{k}\in\{(\texttt{retrieve},\Delta q_{k}),(\texttt{reflect},f_{k}),(\texttt{answer},w_{k})\},} \tag{4} $$ where $\Delta q_{k}$ is a refinement query, $f_{k}$ is a reasoning content, and $w_{k}$ is a draft answer, which are utilized in the downstream action nodes. To ensure stability, router applies three deterministic constraints: 1) a maximum iteration budget $n_{\text{max}}$ that forces an answer action once the budget is exhausted, 2) a reflect-streak capacity $n_{\text{cap}}$ that forces a retrieve action when too many reflections have occurred consecutively, and 3) a retrieval-opportunity check that switches the action to reflect whenever the retrieval stage returns no snippets. The router’s algorithm is shown in Alg. 1. Algorithm 1 Router policy in MemR 3 1: Input: query $q$ , previous snippets $\mathcal{S}_{k-1}$ , iteration $k$ , budgets $n_{\text{max}},n_{\text{cap}}$ , current reflect-streak length $n_{\text{streak}}$ . 2: Output: action $a_{k}$ . 3: if $k≥ n_{\text{max}}$ then 4: $a_{k}=\texttt{answer}$ $\triangleright$ Max iteration budget. 5: else if $\mathcal{S}_{k-1}=\emptyset$ then 6: $a_{k}=\texttt{reflect}$ $\triangleright$ No retrieved snippets. 7: else if $n_{\text{streak}}≥ n_{\text{cap}}$ then 8: $a_{k}=\texttt{retrieve}$ $\triangleright$ Max reflect streak. 9: else 10: pass $\triangleright$ Keep the generated action. 11: end if These lightweight rules stabilize the decision process while preserving flexibility. We further introduce the detailed implementation of these constraints when introducing the system prompt in Appendix A.1. Retrieve. Given a generated refinement $\Delta q_{k}$ , the retrieve node constructs $q_{k}^{\mathrm{ret}}=q\oplus\Delta q_{k}$ , where $\oplus$ means textual combination and $q$ is the original query, and then, fetches new memory snippets: $$ \begin{split}\mathcal{S}_{k}=\texttt{Retrieve}(q_{k}^{\mathrm{ret}},\mathcal{M}\backslash\mathcal{M}^{\text{ret}}_{k-1}),~\mathcal{M}^{\text{ret}}_{k}=\mathcal{M}^{\text{ret}}_{k-1}\cup\mathcal{S}_{k}.\end{split} \tag{5} $$ Snippets $\mathcal{S}_{k}$ are independently used for the next generation without history accumulation. Moreover, retrieved snippets are masked to prevent re-selection. A major benefit of MemR 3 is that it treats all concrete retrievers as plug-in modules. Any retriever, e.g., vector search, graph memory, hybrid stores, or future systems, can be integrated into MemR 3 as long as they return textual snippets, optionally with stable identifiers that can be masked once used. This abstraction ensures MemR 3 remains lightweight, portable, and compatible. Reflect. The reflect node incorporates the reasoning process $\mathcal{F}_{k-1}$ , and invokes the router to update $(\mathcal{E}_{k},\mathcal{G}_{k},a_{k})$ in Eq. 3, where evidence and gaps can be re-summarized. Answer. Once the router selects answer, the final answer is generated from the original query $q$ , the draft answer $w_{k}$ , evidence $\mathcal{E}_{k}$ using prompt $p_{w}$ from rasmussen2025zep: $$ w\leftarrow\texttt{LLM}(q,w_{k},\mathcal{E}_{k},p_{w}), \tag{6} $$ The answer LLM is instructed to avoid hallucinations and remain faithful to evidence. 3.5 Discussion on Efficiency Although MemR 3 introduces extra routing steps, it maintains low overhead via 1) Compact evidence and gap summaries: only short summaries are repeatedly fed into the router. 2) Masked retrieval: each retrieval call yields genuinely new information. 3) Small iteration budgets: typically, most questions can be answered using only a single iteration. Those complicated questions that require multiple iterations are constrained with a small maximum iteration budget. These design choices ensure that MemR 3 improves retrieval quality without large increases in retrieved tokens. 4 Experiments The experiments are conducted on a machine with an AMD EPYC 7713P 64-core processor, an A100-SXM4-80GB GPU, and 512GB of RAM. Each experiment of MemR 3 is repeated three times to report the average scores. Code available: https://github.com/Leagein/memr3. 4.1 Experimental Protocols Datasets. In line with baselines (xu2025amem; chhikara2025mem0), we employ LoCoMo (maharana2024evaluating) dataset as a fundamental benchmark. LoCoMo has a total of 10 conversations across four categories: 1) multi-hop, 2) temporal, 3) open-domain, 4) single-hop, and 5) adversarial. We exclude the last ‘adversarial’ category, following existing work (chhikara2025mem0; wang2025mirix), since it is used to test whether unanswerable questions can be identified. Each conversation has approximately 600 dialogues with 26k tokens and 200 questions on average. Metrics. We adopt the LLM-as-a-Judge (J) score to evaluate answer quality following chhikara2025mem0; wang2025mirix. Compared with surface-level measures such as F1 or BLEU-1 (xu2025amem; 10738994), this metric better avoids relying on simple lexical overlap and instead captures semantic alignment. Specifically, GPT-4.1 (openai2025gpt41) is employed to judge whether the answer is correct according to the original question and the generated answer, following the prompt by chhikara2025mem0. Table 1: LLM-as-a-Judge scores (%, higher is better) for each question category in the LoCoMo (maharana2024evaluating) dataset. The best results using each LLM backend, except Full-Context, are in bold. | LLM GPT-4o-mini LangMem (langmem_blog2025) | Method A-Mem (xu2025amem) 62.23 | 1. Multi-Hop 61.70 23.43 | 2. Temporal 64.49 47.92 | 3. Open-Domain 40.62 71.12 | 4. Single-Hop 76.63 58.10 | Overall 69.06 | | --- | --- | --- | --- | --- | --- | --- | | Mem0 (chhikara2025mem0) | 67.13 | 55.51 | 51.15 | 72.93 | 66.88 | | | Self-RAG (asai2024self) | 69.15 | 64.80 | 34.38 | 88.31 | 76.46 | | | RAG-CoT-RAG | 71.28 | 71.03 | 42.71 | 86.99 | 77.96 | | | Zep (rasmussen2025zep) | 67.38 | 73.83 | 63.54 | 78.67 | 74.62 | | | MemR 3 (ours, Zep backbone) | 69.39 (+2.01) | 73.83 (+0.00) | 67.01 (+3.47) | 80.60 (+1.93) | 76.26 (+1.64) | | | RAG (lewis2020retrieval) | 68.79 | 65.11 | 58.33 | 83.86 | 75.54 | | | MemR 3 (ours, RAG backbone) | 71.39 (+2.60) | 76.22 (+11.11) | 61.11 (+2.78) | 89.44 (+5.58) | 81.55 (+6.01) | | | Full-Context | 72.34 | 58.88 | 59.38 | 86.39 | 76.32 | | | GPT-4.1-mini | A-Mem (xu2025amem) | 71.99 | 74.77 | 58.33 | 79.88 | 76.00 | | LangMem (langmem_blog2025) | 74.47 | 61.06 | 67.71 | 86.92 | 78.05 | | | Mem0 (chhikara2025mem0) | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 | | | Self-RAG (asai2024self) | 75.89 | 75.08 | 54.17 | 90.12 | 82.08 | | | RAG-CoT-RAG | 80.85 | 81.62 | 62.50 | 90.12 | 84.89 | | | Zep (rasmussen2025zep) | 72.34 | 77.26 | 64.58 | 83.49 | 78.94 | | | MemR 3 (ours, Zep backbone) | 77.78 (+5.44) | 77.78 (+0.52) | 69.79 (+5.21) | 84.42 (+0.93) | 80.88 (+1.94) | | | RAG (lewis2020retrieval) | 73.05 | 73.52 | 62.50 | 85.90 | 79.46 | | | MemR 3 (ours, RAG backbone) | 81.20 (+8.15) | 82.14 (+8.62) | 71.53 (+9.03) | 92.17 (+6.27) | 86.75 (+7.29) | | | Full-Context | 86.43 | 86.82 | 71.88 | 93.73 | 89.00 | | Baselines. We select four groups of advanced methods as baselines: 1) memory systems, including A-mem (xu2025amem), LangMem (langmem_blog2025), and Mem0 (chhikara2025mem0); 2) agentic retrievers, like Self-RAG (asai2024self). We also design a RAG-CoT-RAG (RCR) pipeline beyond ReAct (yao2022react) as a strong agentic retriever baseline combining both RAG (lewis2020retrieval) and Chain-of-Thoughts (CoT) (wei2022chain); 3) backend baselines, including chunk-based (RAG (lewis2020retrieval)) and graph-based (Zep (rasmussen2025zep)) memory storage, demonstrating the plug-in capability of MemR 3 across different retriever backends; 4) Moreover, ‘Full-Context’ is widely used as a strong baseline and, when the entire conversation fits within the model window, serves as an empirical upper bound on J score (chhikara2025mem0; wang2025mirix). More detailed introduction of these baselines is shown in Appendix C.1. Other Settings. Other experimental settings and protocols are shown in Appendix C.2. LLM Backend. We reviewed recent work and found that it most frequently used GPT-4o-mini (openai2024gpt4omini), as it is inexpensive and performs well. While some work (wang2025mirix) also includes GPT-4.1-mini (openai2025gpt41), we set both of them as our LLM backends. In our main results, MemR 3 is performed at temperature 0. 4.2 Main Results Overall. Table 1 reports LLM-as-a-Judge (J) scores across four LoCoMo categories. Across both LLM backends and memory backbones, MemR 3 consistently outperforms its underlying retrievers (RAG and Zep) and achieves strong overall J scores. Under GPT-4o-mini, MemR 3 lifts the overall score of Zep from 74.62% to 76.26%, and RAG from 75.54% to 81.55%, with the latter even outperforming the Full-Context baseline (76.32%). With GPT-4.1-mini, we see the same pattern: MemR 3 improves Zep from 78.94% to 80.88% and RAG from 79.46% to 86.75%, making the RAG-backed variant the strongest retrieval-based system and narrowing the gap to Full-Context (89.00%). As expected, methods instantiated with GPT-4.1-mini are consistently stronger than their GPT-4o-mini counterparts. Full-Context also benefits substantially from the stronger LLM, but under GPT-4o-mini it lags behind the best retrieval-based systems, especially on temporal and open-domain questions. Overall, these results indicate that closed-loop retrieval with an explicit evidence–gap state yields gains primarily orthogonal to the choice of LLM or memory backend, and that MemR 3 particularly benefits from backends that expose relatively raw snippets (RAG) rather than heavily compressed structures (Zep). Multi-hop. Multi-hop questions require chaining multiple pieces of evidence and, therefore, directly test our reflective controller. Under GPT-4o-mini, MemR 3 improves both backbones on this category: the multi-hop J score rises from 68.79% to 71.39% on RAG and from 67.38% to 69.39% on Zep, bringing both close to the Full-Context score (72.34%). With GPT-4.1-mini, the gains are more pronounced: MemR 3 boosts RAG from 73.05% to 81.20% and Zep from 72.34% to 77.78%, outperforming all other baselines and approaching the Full-Context upper bound (86.43%). These consistent gains suggest that explicitly tracking evidence and gaps helps the agent coordinate multiple distant memories via iterative retrieval, rather than relying on a single heuristic pass. Temporal. Temporal questions stress the model’s ability to reason about ordering and dating of events over long horizons, where both under- and over-retrieval can be harmful. Here, MemR 3 delivers some of its most considerable relative improvements. For GPT-4o-mini, the temporal J score of RAG jumps from 65.11% to 76.22%, outperforming both the original RAG and the Zep baseline (73.83%), while MemR 3 with a Zep backbone preserves Zep’s strong temporal accuracy (73.83%). Full-Context performs notably worse in this regime (58.88%), indicating that simply supplying all dialogue turns can hinder temporal reasoning under a weaker backbone. With GPT-4.1-mini, MemR 3 again significantly strengthens temporal reasoning: RAG improves from 73.52% to 82.14%, and Zep from 77.26% to 77.78%, making the RAG-backed MemR 3 the best retrieval-based system and closing much of the remaining gap to Full-Context (86.82%). These findings support our design goal that explicitly modeling “what is already known” versus “what is still missing” helps the agent align and compare temporal relations more robustly. Open-Domain. Open-domain questions are less tied to the user’s personal timeline and often require retrieving diverse background knowledge, which makes retrieval harder to trigger and steer. Despite this, MemR 3 consistently improves over its backbones. Under GPT-4o-mini, MemR 3 increases the open-domain J score of RAG from 58.33% to 61.11% and that of Zep from 63.54% to 67.01%, with the Zep-backed variant achieving the best performance among all methods in this block, surpassing Full-Context (59.38%). With GPT-4.1-mini, the gains become even larger: MemR 3 lifts RAG from 62.50% to 71.53% and Zep from 64.58% to 69.79%, nearly matching the Full-Context baseline (71.88%) and again outperforming all other baselines. We attribute these improvements to the router’s ability to interleave retrieval with reflection: when initial evidence is noisy or off-topic, MemR 3 uses the gap representation to reformulate queries and pull in more targeted external knowledge rather than committing to an early, brittle answer. Single-hop. Single-hop questions can often be answered from a single relevant memory snippet, so the potential headroom is smaller, but MemR 3 still yields consistent gains. With GPT-4o-mini, MemR 3 raises the single-hop J score from 78.67% to 80.60% on Zep and from 83.86% to 89.44% on RAG, with the latter surpassing the Full-Context baseline (86.39%). Under GPT-4.1-mini, MemR 3 improves Zep from 83.49% to 84.42% and RAG from 85.90% to 92.17%, making the RAG-backed variant the strongest method overall aside from Full-Context (93.73%). Together with the iteration-count analysis in Sec. 4.3, these results suggest that the router often learns to terminate early on straightforward single-hop queries, gaining accuracy primarily through better evidence selection rather than additional reasoning depth, and thus adding little overhead in tokens or latency. 4.3 Other Experiments We ablate various hyperparameters and modules to evaluate their impact in MemR 3 with the RAG retriever. During these experiments, we utilize GPT-4o-mini as a consistent LLM backend. Table 2: Ablation studies. Best results are in bold. | RAG MemR 3 w/o mask | 68.79 71.39 62.41 | 65.11 76.22 68.54 | 58.33 61.11 55.21 | 83.86 89.44 72.17 | 75.54 81.55 68.54 | | --- | --- | --- | --- | --- | --- | | w/o $\Delta q_{k}$ | 66.67 | 75.08 | 60.42 | 83.37 | 77.11 | | w/o reflect | 65.25 | 73.83 | 61.46 | 83.37 | 76.65 | - MH = Multi-hop; OD = Open-domain; SH = Single-hop. Ablation Studies. We first examine the contribution of the main design choices in MemR 3 by progressively removing them while keeping the RAG retriever and all hyperparameters fixed. As shown in Table 2, disabling masking for previously retrieved snippets (w/o mask) results in the largest degradation, reducing the overall J score from 81.55% to 68.54% and harming every category. This confirms that repeatedly surfacing the same memories wastes budget and fails to effectively close the remaining gaps. Removing the refinement query $\Delta q_{k}$ (w/o $\Delta q_{k}$ ) has a milder effect: temporal and open-domain performance changed a little, but multi-hop and single-hop scores decline significantly, indicating that tailoring retrieval queries from the current evidence-gap state is particularly beneficial for simpler questions. Disabling the reflect node (w/o reflect) similarly reduces performance (from 81.55% to 76.65%), with notable drops on multi-hop and single-hop questions, highlighting the value of interleaving reasoning-only steps with retrieval. Note that in Table 2, the raw retrieved snippets are only visible to the vanilla RAG. Effect of $n_{\text{chk}}$ and $n_{\text{max}}$ . We first choose a nominal configuration for MemR 3 (with a RAG retriever) by arbitrarily setting the number of chunks per iteration $n_{\text{chk}}=3$ and the max iteration budget $n_{\text{max}}=5$ . In Fig. 4(a), we fix $n_{\text{max}}=5$ and perform ablations over $n_{\text{chk}}∈\{1,3,5,7,9\}$ . In Fig. 4(b), we fix $n_{\text{chk}}=3$ and perform ablations over $n_{\text{max}}∈\{1,2,3,4,5\}$ . Considering both of the LLM-as-a-Judge score and token consumption, we eventually choose $n_{\text{chk}}=5$ and $n_{\text{max}}=5$ in all main experiments. <details> <summary>x4.png Details</summary>  ### Visual Description ## Line Chart: LLM-as-a-Judge Performance by Task Category ### Overview This image is a line chart illustrating the performance of an "LLM-as-a-Judge" metric across four distinct task categories as the number of chunks per iteration increases. The chart tracks how increasing the data volume (chunks) affects the judge's accuracy or score across different types of reasoning tasks. ### Components/Axes * **Vertical Axis (Y-axis):** * **Label:** LLM-as-a-Judge (%) * **Scale:** 0 to 100, with major tick marks every 20 units (0, 20, 40, 60, 80). * **Horizontal Axis (X-axis):** * **Label:** # chunks / iteration * **Scale:** Discrete values: 1, 3, 5, 7, 9. * **Legend:** * **Placement:** Located in the center-left to bottom-left region of the plot area, enclosed in a white box with a light border. * **Categories:** * **Multi-hop:** Teal/Dark Green line with solid circular markers. * **Temporal:** Dark Blue line with open square markers. * **Open-domain:** Light Orange/Tan line with open star markers. * **Single-hop:** Magenta/Pink line with solid circular markers. * **Grid:** A light gray dashed grid is present across the background of the plot. ### Detailed Analysis The chart contains four data series. Values below are approximate based on visual alignment with the axes. | # chunks / iteration | Single-hop (Magenta) | Temporal (Blue) | Multi-hop (Teal) | Open-domain (Orange) | | :--- | :--- | :--- | :--- | :--- | | **1** | ~86% | ~66% | ~57% | ~58% | | **3** | ~89% | ~76% | ~69% | ~59% | | **5** | ~89% | ~76% | ~71% | ~61% | | **7** | ~90% | ~76% | ~76% | ~60% | | **9** | ~90% | ~77% | ~78% | ~58% | #### Trend Verification: * **Single-hop (Top-most line):** Slopes upward slightly from 1 to 3 chunks and then plateaus, maintaining the highest performance throughout. * **Temporal (Second from top at start):** Slopes upward sharply from 1 to 3 chunks, then remains almost perfectly flat through 9 chunks. * **Multi-hop (Third from top at start):** Slopes upward steadily across the entire range, eventually crossing the Temporal line at 7 chunks and ending as the second-highest series. * **Open-domain (Bottom-most line):** Remains relatively flat with a very slight peak at 5 chunks before trending slightly downward toward 9 chunks. ### Key Observations * **Performance Ceiling:** The "Single-hop" category consistently outperforms all others, starting at a high baseline (~86%) and reaching a near-ceiling of ~90%. * **Diminishing Returns:** For "Temporal" and "Single-hop" tasks, increasing the number of chunks beyond 3 provides negligible performance gains. * **Growth Potential:** "Multi-hop" tasks show the most significant and sustained improvement as the number of chunks increases, suggesting these tasks benefit more from additional context or data iterations. * **Stagnation:** "Open-domain" performance is largely unaffected by the number of chunks, hovering around the 60% mark. ### Interpretation The data suggests that the complexity of the task determines how much an LLM judge benefits from increased data chunks. * **Single-hop** tasks are likely straightforward enough that the judge reaches peak accuracy with very little context. * **Multi-hop** tasks, which require connecting multiple pieces of information, show a clear positive correlation between data volume and judge performance, indicating that more "chunks" help the model resolve complex dependencies. * **Temporal** tasks seem to require a specific threshold of information (reached at 3 chunks) to be judged effectively, after which additional data does not help. * **Open-domain** tasks may be limited by the model's internal knowledge or the quality of the chunks rather than the quantity, as increasing chunks does not lead to a performance breakthrough. </details> (a) <details> <summary>x5.png Details</summary>  ### Visual Description ## Line Chart: LLM-as-a-Judge Performance by Category ### Overview This image is a line chart illustrating the performance of a Large Language Model (LLM) across four distinct task categories as the number of "max iterations" increases. The performance is measured as a percentage using an "LLM-as-a-Judge" metric. The chart shows that while performance varies significantly between categories, it remains relatively stable or shows slight improvement as iterations increase from 1 to 5. ### Components/Axes * **Vertical Axis (Y-axis):** * **Label:** "LLM-as-a-Judge (%)" * **Scale:** 0 to 100, with major tick marks and grid lines at intervals of 20 (0, 20, 40, 60, 80). * **Horizontal Axis (X-axis):** * **Label:** "max iterations" * **Scale:** 1 to 5, with integer markers at 1, 2, 3, 4, and 5. * **Legend:** * **Title:** "Categories" * **Placement:** Centered vertically on the left side of the plot area, partially overlapping the data lines. * **Items:** * **Multi-hop:** Teal/Dark Green line with solid circular markers. * **Temporal:** Dark Blue/Purple line with open square markers. * **Open-domain:** Light Orange/Peach line with open diamond/star markers. * **Single-hop:** Pink/Magenta line with solid circular markers. ### Content Details The chart tracks four data series across five iteration points. | Category | Iteration 1 | Iteration 2 | Iteration 3 | Iteration 4 | Iteration 5 | Visual Trend | | :--- | :---: | :---: | :---: | :---: | :---: | :--- | | **Single-hop** (Pink) | ~86% | ~87% | ~88% | ~88% | ~89% | Stable, slight upward slope. | | **Temporal** (Dark Blue) | ~71% | ~72% | ~71% | ~75% | ~76% | Upward trend with a minor dip at iteration 3. | | **Multi-hop** (Teal) | ~68% | ~69% | ~68% | ~68% | ~69% | Very stable/flat. | | **Open-domain** (Orange) | ~58% | ~60% | ~59% | ~60% | ~60% | Stable, slight initial rise then flat. | ### Key Observations * **Performance Hierarchy:** There is a clear and consistent performance hierarchy across all iteration counts: Single-hop > Temporal > Multi-hop > Open-domain. * **Stability:** Most categories show very little change in performance as iterations increase. The lines are largely horizontal. * **Temporal Improvement:** The "Temporal" category shows the most significant relative improvement, gaining approximately 5 percentage points from iteration 1 to iteration 5. * **Saturation:** Performance appears to plateau quickly. For most categories, there is no significant gain after 2 iterations. ### Interpretation * **Task Complexity:** The data suggests that "Single-hop" tasks are the least complex for the model to handle, achieving high accuracy (~85-90%) almost immediately. Conversely, "Open-domain" tasks represent the highest complexity or highest uncertainty, with performance struggling to break the 60% threshold. * **Iterative Gains:** The "max iterations" parameter likely refers to a reasoning or refinement process (like Chain-of-Thought or self-correction). The fact that performance does not increase dramatically suggests that the model either reaches its correct conclusion early or that additional iterations do not provide enough new information to correct initial errors. * **Temporal Reasoning:** The slight upward trend in "Temporal" tasks suggests that these specific types of problems benefit more from iterative processing than static "Multi-hop" or "Open-domain" tasks, perhaps because temporal sequences require more steps of logical verification. * **LLM-as-a-Judge Reliability:** Since the metric is "LLM-as-a-Judge," the results reflect the model's ability to satisfy a set of evaluative criteria. The high scores in "Single-hop" might indicate that these tasks are easier to evaluate objectively compared to the potentially more subjective or vast "Open-domain" tasks. </details> (b) Figure 4: LLM-as-a-Judge score (%) with different a) number of chunks per iteration and b) max iterations. Iteration count. We further inspect how often MemR 3 actually uses multiple retrieve/reflect/answer iterations when $n_{\text{chk}}=5$ and $n_{\text{max}}=5$ (Fig. 5). Overall, most questions are answered after a single iteration, and this effect is particularly strong for Single-hop questions. An exception is open-domain questions, for which 58 of 96 require continuous retrieval or reflection until the maximum number of iterations is reached, highlighting the inherent challenges and uncertainty in these questions. Additionally, only a small fraction of questions terminate at intermediate depths (2–4 iterations), suggesting that MemR 3 either becomes confident early or uses the whole iteration budget when the gap remains non-empty. We observe that this distribution arises from two regimes. On the one hand, straightforward questions require only a single piece of evidence and can be resolved in a single iteration, consistent with intuition. From the perspective of the idealized tracker in Appendix B, these are precisely the queries for which every requirement $r∈ R(q)$ is supported by some retrieved memory item $m∈\bigcup_{j≤ k}S_{j}$ with $m\models r$ , so the completeness condition in Theorem B.4 is satisfied and the ideal gap $G_{k}^{\star}$ becomes empty. On the other hand, some challenging questions are inherently underspecified given the stored memories, so the gap cannot be fully closed even if the agent continues to refine its query. For example, for the question “ When did Melanie paint a sunrise? ”, the correct answer in our setup is simply “ 2022 ” (the year). MemR 3 quickly finds this year at the first iteration based on evidence “ Melanie painted the lake sunrise image last year (2022). ”. However, under the idealized abstraction, the requirement set $R(q)$ implicitly includes an exact date predicate (year–month–day), and no memory item $m∈\bigcup_{j≤ K}S_{j}$ satisfies $m\models r$ for that finer-grained requirement. Thus, the precondition of Theorem B.4 (3) is violated, and $G_{k}^{\star}$ never becomes empty; the practical tracker mirrors this by continuing to search for the missing specificity until it hits the maximum iteration budget. In such cases, the additional token consumption is primarily due to a mismatch between the question’s granularity and the available memory, rather than a failure of the agent. <details> <summary>x6.png Details</summary>  ### Visual Description ## Bar Chart: Iteration Count by Category and Iteration Number ### Overview This image is a grouped bar chart illustrating the distribution of "Iteration Count" across four distinct "Categories" (Multi-hop, Temporal, Open-Domain, and Single-hop). The data is segmented into five sequential iterations, represented by different shades of teal and blue. A notable feature of this chart is a broken Y-axis to accommodate a significant outlier in the "Single-hop" category. ### Components/Axes * **Y-Axis (Vertical):** * **Label:** "Iteration Count" (positioned center-left). * **Scale:** The axis starts at 0 and increments by 50 up to 200. There is a visual break (indicated by two parallel dashed lines with "zigzag" break symbols) between 200 and 650. The scale then resumes with increments of 25 (650, 675, 700). * **X-Axis (Horizontal):** * **Label:** "Categories" (positioned bottom-center). * **Categories:** Multi-hop, Temporal, Open-Domain, Single-hop. * **Legend:** * **Title:** "# Iterations" * **Position:** Top-left, inside the chart area. * **Items:** * **iteration 1:** Dark teal (leftmost bar in each group). * **iteration 2:** Medium teal. * **iteration 3:** Light teal. * **iteration 4:** Light blue. * **iteration 5:** Very light blue (rightmost bar in each group). ### Content Details The following table reconstructs the data points displayed above each bar in the chart. | Category | Iteration 1 | Iteration 2 | Iteration 3 | Iteration 4 | Iteration 5 | | :--- | :---: | :---: | :---: | :---: | :---: | | **Multi-hop** | 184 | 1 | 2 | 0 | 95 | | **Temporal** | 168 | 4 | 1 | 0 | 148 | | **Open-Domain** | 36 | 2 | 0 | 0 | 58 | | **Single-hop** | 690 | 1 | 1 | 1 | 137 | #### Visual Trend Verification: * **Iteration 1 (Dark Teal):** Shows the highest count in three out of four categories. It exhibits a massive spike in the "Single-hop" category. * **Iterations 2, 3, and 4:** These series consistently show near-zero values across all categories, appearing as flat lines or very tiny bars at the baseline. * **Iteration 5 (Very Light Blue):** Shows a significant "rebound" or secondary peak in all categories. In the "Open-Domain" category, it is actually higher than Iteration 1. ### Key Observations * **The "Single-hop" Outlier:** The Iteration 1 count for "Single-hop" (690) is more than triple the next highest value in that iteration (184 for Multi-hop). This necessitates the Y-axis break. * **Bimodal Distribution:** Across all categories, there is a distinct "U-shaped" or bimodal pattern where Iteration 1 and Iteration 5 contain almost all the data, while the middle iterations (2-4) are nearly empty. * **Open-Domain Exception:** This is the only category where the final iteration (58) exceeds the initial iteration (36). ### Interpretation The data suggests a process—likely an algorithmic search, reasoning chain, or data processing pipeline—that typically resolves or reaches a state of interest either immediately (Iteration 1) or after a fixed number of steps (Iteration 5). * **Efficiency/Directness:** The high values in Iteration 1 for "Single-hop" (690) compared to "Multi-hop" (184) suggest that the system is much more likely to find a result in a single step when the task is simple (single-hop). * **The "Dead Zone":** The near-zero values for iterations 2, 3, and 4 indicate that if a result isn't found immediately, it is rarely found in the intermediate steps. * **The "Final Effort":** The resurgence in Iteration 5 suggests a "fallback" mechanism or a concluding processing step that triggers after the intermediate steps fail to produce a result. In "Open-Domain" tasks, which are inherently broader and potentially more difficult, this final step is more productive than the initial attempt, suggesting that the complexity of the category requires more "thinking" or "hops" to resolve. </details> Figure 5: Number of questions requiring different numbers of iterations before final answers, across four categories. 4.4 Revisiting the Evaluation Protocols of LoCoMo During our reproduction of the baselines, we identified a latent ambiguity in the LoCoMo dataset’s category indexing. Specifically, the mapping between numerical IDs and semantic categories (e.g., Multi-hop vs. Single-hop) implies a non-trivial alignment challenge. We observed that this ambiguity has led to category misalignment in several recent studies (chhikara2025mem0; wang2025mirix), potentially skewing the granular analysis of agent capabilities. To ensure a rigorous and fair comparison, we recalibrate the evaluation protocols for all baselines. In Table 1, we report the performance based on the corrected alignment, where the alignment can be induced by the number of questions in each category. We believe this clarification contributes to a more accurate understanding of the current SOTA landscape. Details of the dataset realignment are illustrated in Appendix C.3. 5 Conclusion In this work, we introduce MemR 3, an autonomous memory-retrieval controller that transforms standard retrieve-then-answer pipelines into a closed-loop process via a LangGraph-based sequential decision-making framework. By explicitly maintaining what is known and what remains unknown using an evidence-gap tracker, MemR 3 can iteratively refine queries, balance retrieval and reflection, and terminate early once sufficient evidence has been gathered. Our experiments on the LoCoMo benchmark show that MemR 3 consistently improves LLM-as-a-Judge scores over strong memory baselines, while incurring only modest token and latency overhead and remaining compatible with heterogeneous backends. Beyond these concrete gains, MemR 3 offers an explainable abstraction for reasoning under partial observability in long-horizon agent settings. However, we acknowledge some limitations for future work: 1) MemR 3 requires an existing retriever or memory structure, and particularly, the performance greatly depends on the retriever or memory structure. 2) The routing structure could lead to token waste for answering simple questions. 3) MemR 3 is currently not designed for multi-modal memories like images or audio. Appendix A Prompts A.1 System prompt of the generate node The system prompt is defined as follows, where the “decision_directive” instructs the maximum iteration budges, reflect-streak capacity, and retrieval opportunity check, introduced in Sec. 3.4. Generally, “decision_directive” is a textual instruction: “reflect” if you need to think about the evidence and gaps; choose “answer” ONLY when evidence is solid and no gaps are noted; choose “retrieve” otherwise. However, when the maximum iterations budget is reached, “decision_directive” is set as “answer” to stop early. When the reflection reaches the maximum capacity, “decision_directive” is set as “retrieve” to avoid repeated ineffective reflection. When there is no useful retrieval remains, “decision_directive” is set as “reflect” to avoid repeated ineffective retrieval. Through these constraints, the agent can avoid infinite ineffective actions to maintain stability. System Prompt You are a memory agent that plans how to gather evidence before producing the final response shown to the user. Always reply with a strict JSON object using this schema: - evidence: JSON array of concise factual bullet strings relevant to the user’s question; preserve key numbers/names/time references. If exact values are unavailable, include the most specific verified information (year/range) without speculation. Never mention missing or absent information here – “gaps” will do that. - gaps: gaps between the question and evidence that prevent a complete answer. - decision: one of [“retrieve”,“answer”,“reflect”]. Choose {decision_directive}. Only include these conditional keys: - retrieval_query: only when decision == “retrieve”. Provide a STANDALONE search string; short (5-15 tokens). * BAD Query: “the date” (lacks context). * GOOD Query: “graduation ceremony date” (specific). * STRATEGY: 1. Search for the ANCHOR EVENT. (e.g. Question: “What happened 2 days after X?”, Query: “timestamp of event X”). 2. Search for the MAPPED ENTITY. (e.g. Question: “Weather in the Windy City”, Query: “weather in Chicago”). - detailed_answer: only when decision == “answer”; response using current evidence (keep absolute dates, avoid speculation). If evidence is limited, provide only what is known, or make cautious inferences grounded solely in that limited evidence. Do not mention missing or absent information in this field. - reasoning: only when decision == “reflect”; if further retrieval is unlikely, use current evidence to think step by step through the evidence and gaps, and work toward the answer, including any time normalization. Never include extra keys or any text outside the JSON object. A.2 User prompt of the generate node Apart from the system, the user prompt is responsible to feed additional information to the LLM. Specifically, at the $k$ iteration, “question” is the original question $q$ . “evidence_block” and “gap_block” are evidence $\mathcal{E}_{k}$ and gaps $\mathcal{G}_{k}$ introduced in Sec. 3.3. “raw_block” is the retrieved raw snippets $\mathcal{S}_{k}$ in Eq. 5. “reasoning_block” is the reasoning content $\mathcal{F}_{k}$ in Sec. 3.4. “last_query” is the refined query $\Delta q_{k}$ introduced in Sec. 3.4 that enables the new query to be different from the prior one. Note that these fields can be left empty if the corresponding information is not present. User Prompt # Question {question} # Evidence {evidence_block} # Gaps {gap_block} # Memory snippets {raw_block} # Reasoning {reasoning_block} # Prior Query {last_query} # INSTRUCTIONS: 1. Update the evidence as a JSON ARRAY of concise factual bullets that directly help answer the question (preserve key numbers/names/time references; use the most specific verified detail without speculation). 2. Update gaps: remove resolved items, add new missing specifics blocking a full answer, and set to “None” when nothing is missing. 3. If you produce a retrieval_query, make sure it differs from the previous query. 4. Decide the next action and return ONLY the JSON object described in the system prompt. Appendix B Formalizing the Evidence-Gap Tracker A central component of MemR 3 is the evidence-gap tracker introduced in Sec. 3.3, which maintains an evolving summary of i) what information has been reliably established from memory and ii) what information is still missing to answer the query. While the practical implementation of this tracker is based on LLM-generated summaries, we introduce an idealized formal abstraction that clarifies its intended behavior, enables principled analysis, and provides a foundation for studying correctness and robustness. This abstraction does not assume perfect extraction; rather, the LLM acts as a stochastic approximator to the idealized tracker. **Definition B.1 (Idealized Requirement Space)** *For a user query $q$ , we define a finite set of atomic information requirements, which specify the minimal facts needed to fully answer the query: $$ R(q)=\{r_{1},r_{2},\dots,r_{m}\}. \tag{7} $$* For example, for the question “How many months passed between events $A$ and $B$ ?”, the requirement set can be $$ R(q)=\{\text{date}(A),\text{date}(B)\}. \tag{8} $$ Each requirement $r∈ R(q)$ is associated with a symbolic predicate (e.g., a timestamp, entity attribute, or event relation), and $R(q)$ provides the semantic target against which retrieved memories are judged. **Definition B.2 (Memory-Support Relation)** *Let $\mathcal{M}$ be the memory store and $S_{k}⊂eq\mathcal{M}$ denote the snippets retrieved at iteration $k$ . We define a relation $m\models r$ to indicate that memory item $m∈\mathcal{M}$ contains sufficient information to support requirement $r∈ R(q)$ . Formally, $m\models r$ holds if the textual content of $m$ contains a minimal witness (e.g., a timestamp, entity mention, or explicit assertion) matching the predicate corresponding to $r$ . The matching criterion may be implemented via deterministic pattern rules or LLM-based semantic matching; our analysis is agnostic to this choice.* **Definition B.3 (Idealized Evidence-Gap Update Rule)** *At iteration $k$ , the idealized tracker maintains two sets: i) the evidence $E_{k}⊂eq R(q)$ and ii) the gaps $G_{k}=R(q)\setminus E_{k}$ . Given newly retrieved snippets $S_{k}$ , the ideal updates are $$ E_{k}^{\star}=E_{k-1}\cup\big\{r\in R(q)\,\big|\,\exists m\in S_{k},\;m\models r\big\},\qquad G_{k}^{\star}=R(q)\setminus E_{k}^{\star}. \tag{9} $$* In this abstraction, the tracker monotonically accumulates verified requirements and removes corresponding gaps, providing a clean characterization of the desired system behavior independent of noise. B.1 Practical Instantiation via LLM Summaries In MemR 3, the tracker is instantiated through LLM-generated summaries: $$ (E_{k},G_{k})=\mathrm{LLM}\big(q,S_{k},E_{k-1},G_{k-1}\big), \tag{10} $$ where the prompt explicitly instructs the model to: (i) extract concise factual bullets relevant to $q$ , (ii) enumerate missing information blocking a complete answer, and (iii) avoid hallucinations or speculative inference. Thus, $(E_{k},G_{k})$ serves as a stochastic approximation to the idealized $(E_{k}^{\star},G_{k}^{\star})$ : $$ (E_{k},G_{k})\approx(E_{k}^{\star},G_{k}^{\star}), \tag{11} $$ with deviations arising from LLM extraction noise. This perspective reconciles the formal update rule with the prompt-driven practical implementation. B.2 Correctness Properties under Idealized Extraction Although the practical instantiation lacks deterministic guarantees, the idealized tracker in Definition B.3 satisfies several intuitive properties essential for closed-loop retrieval. **Theorem B.4 (Properties of the Idealized Tracker)** *Assume that for all $k$ and all $r∈ R(q)$ , we have $r∈ E_{k}^{\star}$ if and only if there exists some $m∈\bigcup_{j≤ k}S_{j}$ such that $m\models r$ . Then the following hold: 1. Monotonicity: $E_{k-1}^{\star}⊂eq E_{k}^{\star}$ and $G_{k}^{\star}⊂eq G_{k-1}^{\star}$ for all $k≥ 1$ . 1. Soundness: If $m\models r$ for some retrieved memory $m∈ S_{k}$ , then $r∈ E_{k}^{\star}$ . 1. Completeness at convergence: If every requirement $r∈ R(q)$ is supported by some $m∈\bigcup_{j≤ K}S_{j}$ with $m\models r$ , then $E_{K}^{\star}=R(q)$ and hence $G_{K}^{\star}=\varnothing$ .* * Proof* (1) By Definition B.3, $$ E_{k}^{\star}=E_{k-1}^{\star}\cup\big\{r\in R(q)\,\big|\,\exists m\in S_{k},\;m\models r\big\}, \tag{12} $$ so $E_{k-1}^{\star}⊂eq E_{k}^{\star}$ . Since $G_{k}^{\star}=R(q)\setminus E_{k}^{\star}$ and $E_{k-1}^{\star}⊂eq E_{k}^{\star}$ , we obtain $G_{k}^{\star}⊂eq G_{k-1}^{\star}$ . (2) If $m\models r$ for some $m∈ S_{k}$ , then by Definition B.3 we have $r∈\{r^{\prime}∈ R(q)\mid∃ m^{\prime}∈ S_{k},\;m^{\prime}\models r^{\prime}\}⊂eq E_{k}^{\star}$ . (3) If every $r∈ R(q)$ is supported by some $m∈\bigcup_{j≤ K}S_{j}$ with $m\models r$ , then repeated application of the update rule ensures that each such $r$ is eventually added to $E_{K}^{\star}$ . Hence $E_{K}^{\star}=R(q)$ and therefore $G_{K}^{\star}=R(q)\setminus E_{K}^{\star}=\varnothing$ . ∎ These properties characterize the target behavior that the LLM-based tracker implementation aims to approximate. B.3 Robustness Considerations Since real LLMs introduce extraction noise, the practical tracker may deviate from the idealized $(E_{k}^{\star},G_{k}^{\star})$ , for example, through false negatives (missing evidence), false positives (hallucinated evidence), or unstable gap estimates. In the main text (Sec. 3.3 and Sec. 4.3), we study these effects empirically by injecting noisy or contradictory memories and measuring their impact on routing decisions and final answer quality. The formal abstraction above serves as the reference model against which these robustness behaviors are interpreted. B.4 Approximation Bias of the LLM Tracker The abstraction in this section assumes access to an ideal tracker that updates ( $\mathcal{E}_{k}$ , $\mathcal{G}_{k}$ ) exactly according to the requirement–support relation $m\models r$ . In practice, MemR 3 uses an LLM-generated tracker ( $\mathcal{E}_{k}$ , $\mathcal{G}_{k}$ ), which only approximates this ideal update. This introduces several forms of approximation bias: i) Coverage bias (false negatives): supported requirements $r∈ R(q)$ that are omitted from $\mathcal{E}_{k}$ ; ii) Hallucination bias (false positives): requirements $r$ that appear in $\mathcal{E}_{k}$ even though no retrieved memory item supports them; iii) Granularity bias: cases where the tracker records a coarser fact (e.g., a year) but the requirement space $R(q)$ contains a finer predicate (e.g., an exact date), so the ideal requirement is never fully satisfied. B.5 Toy example of the granularity bias The “ Melanie painted a sunrise ” case in Sec. 4.3 provides a concrete illustration of granularity bias. The question asks “ When did Melanie paint a sunrise? ”, and in our setup the correct answer is the year 2022. Under the ideal abstraction, however, the requirement space $R(q)$ implicitly contains a fine-grained predicate $r_{\text{date}}$ corresponding to the full year–month–day of the painting event. The memory store only contains a coarse statement such as “ Melanie painted the lake sunrise image last year (2022). ” In the ideal tracker, no memory item $m$ satisfies $m\models r_{\text{date}}$ , so the precondition of Theorem B.4 ’s completeness clause is violated and the ideal gap $\mathcal{G}_{k}$ never becomes empty. The practical LLM tracker mirrors this behavior: it quickly recovers the year 2022 as evidence, but continues to treat the exact date as a remaining gap, eventually hitting the iteration budget without fully closing Gk. This example shows that some apparent “failures” of the approximate tracker are in fact structural: they arise from a mismatch between the granularity of $R(q)$ and the information actually present in the memory store. Appendix C Experimental Settings C.1 Baselines We select four groups of advanced methods as baselines: 1) memory systems, including A-mem (xu2025amem), LangMem (langmem_blog2025), and Mem0 (chhikara2025mem0); 2) agentic retrievers, like Self-RAG (asai2024self). We also design a RAG-CoT-RAG (RCR) pipeline as a strong agentic retriever baseline combining both RAG (lewis2020retrieval) and Chain-of-Thoughts (CoT) (wei2022chain); 3) backend baselines, including chunk-based (RAG (lewis2020retrieval)) and graph-based (Zep (rasmussen2025zep)) memory storage, demonstrating the plug-in capability of MemR 3 across different retriever backends. Moreover, ‘Full-Context’ is widely used as a strong baseline and, when the entire conversation fits within the model window, serves as an empirical upper bound on J score (chhikara2025mem0; wang2025mirix). More detailed introduction of these baselines is shown in Appendix C.1. We divide our groups into four groups: memory systems, agentic retrievers, backend baselines, and full-context. C.1.1 Memory systems In this group, we consider recent advanced memory systems, including A-mem (xu2025amem), LangMem (langmem_blog2025), and Mem0 (chhikara2025mem0), to demonstrate the comprehensively strong capability of MemR 3 from a memory control perspective. A-mem (xu2025amem) https://github.com/WujiangXu/A-mem. A-Mem is an agent memory module that turns interactions into atomic notes and links them into a Zettelkasten-style graph using embeddings plus LLM-based linking. LangMem (langmem_blog2025). LangMem is LangChain’s persistent memory layer that extracts key facts from dialogues and stores them in a vector store (e.g., FAISS/Chroma) for later retrieval. Mem0 (chhikara2025mem0) https://github.com/mem0ai/mem0. Mem0 is an open-source memory system that enables an LLM to incrementally summarize, deduplicate, and store factual snippets, with an optional graph-based memory extension. C.1.2 Agentic Retrievers In this group, we examine the agentic structures underlying memory retrieval to show the advanced performance of MemR 3 on memory retrieval, and particularly, showing the advantage of the agentic structure of MemR 3. To validate this, we include Self-RAG (asai2024self) and design a strong heuristic baseline, RAG-CoT-RAG (RCR), which combines RAG and CoT (wei2022chain). Self-RAG (asai2024self). A model-driven retrieval controller where the LLM decides, at each step, whether to answer or issue a refined retrieval query. Unlike MemR 3, retrieval decisions in Self-RAG are implicit in the model’s chain-of-thought, without explicit state tracking. We reproduce their original code and prompt to suit our task. RAG-CoT-RAG (RCR). We design a strong heuristic baseline that extends beyond ReAct (yao2022react) by performing one initial retrieval (lewis2020retrieval), a CoT (wei2022chain) step to identify missing information, and a second retrieval using a refined query. It provides multi-step retrieval but lacks an explicit evidence-gap state or a general controller. C.1.3 Backend Baselines In this group, we incorporate vanilla RAG (lewis2020retrieval) and Zep (rasmussen2025zep) as retriever backends for MemR 3 to demonstrate the advantages of MemR 3 ’s plug-in design. The former is a chunk-based method while the latter is a graph-based one, which cover most types of existing memory systems. Vanilla RAG (lewis2020retrieval). The vanilla RAG retrieves the top- $k$ relevant snippets from the query once and provides a direct answer, without iterative retrieval or reasoning-based refinement. The other retrieval setting ( $n_{\text{chk}}$ , chunk size, etc.) is the same as that in MemR 3. Zep (rasmussen2025zep). Zep is a hosted memory service that builds a time-aware knowledge graph over conversations and metadata to support fast semantic and temporal queries. We implement their original code. C.1.4 Full-Context Lastly, we include Full-Context as a strong baseline, which provides the model with the entire conversation or memory buffer without retrieval, serving as an upper-bound reference that is unconstrained by retrieval errors or missing information. C.2 Other protocols. For all chunk-based methods like RAG (lewis2020retrieval), Self-RAG (asai2024self), RAG-CoT-RAG, and MemR 3 (RAG retriever), we set the embedding model as text-embedding-large-3 (openai2024embeddinglarge3) and use a re-ranking strategy (reimers2019sentence) (ms-marco-MiniLM-L-12-v2) to search relevant memories rather than just similar ones. The chunk size is selected from {128, 256, 512, 1024} using the GPT-4o-mini backend when $n_{\text{max}}=1$ and $n_{\text{chk}}=1$ , and we ultimately choose 256. This chunk size is also in line with Mem0 (chhikara2025mem0). Table 3: The alignment of the orders and categories in LoCoMo dataset. | Category Order # Questions | Multi-Hop Category 1 282 | Temporal Category 2 321 | Open-Domain Category 3 96 | Single-Hop Category 4 830 | Adversarial Category 5 445 | | --- | --- | --- | --- | --- | --- | C.3 Re-alignment of LoCoMo dataset Misalignment in existing works. Although the correct order of the different categories is not explicitly reported in LoCoMo (maharana2024evaluating), we can infer it from the number of questions in each category. The correct alignment is shown in Table 3. We believe this clarification could benefit the LLM memory community. Repeated questions in LoCoMo dataset. Note that the number of single-hop and adversarial questions is 841 and 446 in the original LoCoMo, while the number is 830 and 445 based on our count, due to 12 repeated questions. In the following, the first question is repeated in both the single-hop and adversarial categories in the 2rd conversation (we remove the one in the adversarial category), while the remaining 11 questions are repeated in the single-hop category in the 8th conversation. 1. What did Gina receive from a dance contest? (conversation 2, question 62), (conversation 2, question 96) 1. What are the names of Jolene’s snakes? (conversation 8, question 17), (conversation 8, question 90) 1. What are Jolene’s favorite books? (conversation 8, question 26), (conversation 8, question 91) 1. What music pieces does Deborah listen to during her yoga practice? (conversation 8, question 43), (conversation 8, question 92) 1. What games does Jolene recommend for Deborah? (conversation 8, question 59), (conversation 8, question 93) 1. What projects is Jolene planning for next year? (conversation 8, question 62), (conversation 8, question 94) 1. Where did Deborah get her cats? (conversation 8, question 63), (conversation 8, question 95) 1. How old are Deborah’s cats? (conversation 8, question 64), (conversation 8, question 96) 1. What was Jolene doing with her partner in Rio de Janeiro? (conversation 8, question 68), (conversation 8, question 97) 1. Have Deborah and Jolene been to Rio de Janeiro? (conversation 8, question 70), (conversation 8, question 98) 1. When did Jolene’s parents give her first console? (conversation 8, question 73), (conversation 8, question 99) 1. What do Deborah and Jolene plan to try when they meet in a new cafe? (conversation 8, question 75), (conversation 8, question 100) Table 4: Repeated experiments of MemR 3 in the main results in Table 1. | GPT-4o-mini 2 3 | Zep 69.86 70.21 | 1 72.59 75.39 | 68.09 67.71 64.58 | 73.52 80.36 80.72 | 68.75 76.00 76.65 | 80.72 | 76.13 | | --- | --- | --- | --- | --- | --- | --- | --- | | mean $±$ std | 69.39 $±$ 0.41 | 73.83 $±$ 1.40 | 67.01 $±$ 1.64 | 80.60 $±$ 0.18 | 76.26 $±$ 0.33 | | | | RAG | 1 | 71.63 | 77.26 | 61.46 | 89.28 | 81.75 | | | 2 | 70.21 | 76.01 | 59.38 | 89.40 | 81.16 | | | | 3 | 72.34 | 75.39 | 62.50 | 89.64 | 81.75 | | | | mean $±$ std | 71.39 $±$ 1.08 | 76.22 $±$ 0.95 | 61.11 $±$ 1.59 | 89.44 $±$ 0.18 | 81.56 $±$ 0.34 | | | | GPT-4.1-mini | Zep | 1 | 78.72 | 78.50 | 72.92 | 84.34 | 81.36 | | 2 | 75.89 | 77.26 | 68.75 | 84.58 | 80.44 | | | | 3 | 78.72 | 77.57 | 67.71 | 84.34 | 80.84 | | | | mean $±$ std | 77.78 $±$ 1.44 | 77.78 $±$ 0.26 | 69.79 $±$ 1.04 | 84.42 $±$ 0.12 | 80.88 $±$ 0.24 | | | | RAG | 1 | 81.56 | 83.18 | 69.79 | 91.93 | 86.79 | | | 2 | 82.62 | 80.69 | 75.00 | 92.65 | 87.18 | | | | 3 | 79.43 | 82.55 | 69.79 | 91.93 | 86.27 | | | | mean $±$ std | 81.20 $±$ 1.62 | 82.14 $±$ 1.29 | 71.53 $±$ 3.01 | 92.17 $±$ 0.42 | 86.75 $±$ 0.46 | | | Appendix D Experimental Results D.1 Repeated Experiments. For the LoCoMo dataset, we show the repeated experiments of MemR 3 in Table 4. <details> <summary>x7.png Details</summary>  ### Visual Description ## Chart Type: Dual-Axis Bar and Line Chart ### Overview This technical chart compares three different information retrieval/processing methods (**RAG**, **MemR3**, and **Full-Context**) across four distinct task categories. It evaluates these methods based on two metrics: **Retrieved Token Usage** (efficiency) and **LLM-as-a-Judge Score** (performance). The chart uses a logarithmic scale for token usage and a linear percentage scale for the performance score. ### Components/Axes * **X-Axis (Methods):** Categorical axis listing three methods: * **RAG** (Retrieval-Augmented Generation) * **MemR3** * **Full-Context** * **Left Y-Axis (Retrieved Token Usage):** Logarithmic scale ranging from $10^0$ (1) to $10^5$ (100,000). This axis corresponds to the **bar charts**. * **Right Y-Axis (LLM-as-a-Judge Score %):** Linear scale ranging from 50% to 100%. This axis corresponds to the **line graphs**. * **Legends:** * **Top-Left (Categories):** Defines the color-coding and markers for both bars and lines: * **Teal Circle ($\bullet$):** Multi-hop * **Dark Blue Square ($\blacksquare$):** Temporal * **Tan Star ($\star$):** Open-domain * **Pink Hexagon ($\hexagon$):** Single-hop * **Top-Center/Right (Data Types):** * **Grey Bar:** Token Usage (bar) * **Black Line with Circle:** LLM-as-a-Judge (line) ### Content Details #### 1. Retrieved Token Usage (Bars - Left Y-Axis) The bars represent the volume of data processed. * **RAG:** All four categories use a uniform amount of tokens, approximately **$1.2 \times 10^3$ (1,200 tokens)**. * **MemR3:** Shows slight variation by category, averaging around **$2 \times 10^3$ (2,000 tokens)**. * Multi-hop: ~1,800 * Temporal: ~2,000 * Open-domain: ~2,200 * Single-hop: ~1,500 * **Full-Context:** All four categories use a uniform, significantly higher amount of tokens, approximately **$2.5 \times 10^4$ (25,000 tokens)**. #### 2. LLM-as-a-Judge Score (Lines - Right Y-Axis) The lines track performance across the methods. * **Multi-hop (Teal):** Slopes upward from RAG (~69%) to MemR3 (~72%) and reaches its peak at Full-Context (~73%). * **Temporal (Dark Blue):** Shows a sharp upward trend from RAG (~65%) to a peak at MemR3 (~78%), followed by a significant drop at Full-Context (~58%). * **Open-domain (Tan):** Remains relatively flat with a very slight peak at MemR3 (~60%) compared to RAG (~58%) and Full-Context (~59%). * **Single-hop (Pink):** The highest performing category. It slopes upward from RAG (~84%) to a peak at MemR3 (~88%), then dips slightly at Full-Context (~86%). ### Key Observations * **Token Efficiency:** Full-Context uses roughly **10x to 20x more tokens** than MemR3 and RAG, representing a massive increase in computational cost. * **Performance Peak:** For three out of four categories (Temporal, Open-domain, Single-hop), **MemR3 achieves the highest performance score**, despite using significantly fewer tokens than Full-Context. * **The "Temporal" Anomaly:** The Temporal category shows a unique pattern where Full-Context performance (~58%) is actually worse than the baseline RAG (~65%) and much worse than MemR3 (~78%). * **Task Difficulty:** "Single-hop" tasks are consistently the easiest for all models (80%+ scores), while "Open-domain" tasks are the most challenging (all scores $\leq$ 60%). ### Interpretation The data suggests that **MemR3 is the most optimized method** among those tested. It provides a "sweet spot" of high performance and low token usage. The fact that MemR3 outperforms Full-Context in most categories—especially the dramatic 20-point lead in "Temporal" tasks—indicates that providing an LLM with the entire context (Full-Context) can introduce noise or "distractions" that degrade performance compared to a more targeted retrieval or memory management system like MemR3. The "Temporal" drop-off in Full-Context is a classic example of the "lost in the middle" phenomenon or context-window saturation, where the model struggles to maintain chronological or sequential logic when overwhelmed by a massive, unfiltered context. Conversely, MemR3's success suggests it effectively filters or structures information to maintain these temporal relationships. </details> Figure 6: Average token consumption of the retrieved snippets (left y-axis) and LLM-as-a-Judge (J) Score (right y-axis) of RAG, MemR 3, and Full-Context across four categories. D.2 Token Consumption In Table 6, we compare the average token consumption of the retrieved snippets and J score of RAG, MemR 3, and Full-Context methods across four categories. The chunk size of RAG and MemR 3 are both set as $n_{\text{chk}}=5$ , while $n_{\text{max}}=2$ for MemR 3. We observe that MemR 3 outperforms RAG across all four categories with only a few additional tokens. While Full-Context consumes significantly more tokens than MemR 3, it surpasses MemR 3 only on multi-hop questions.