# 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
\n
## Diagram: LLM Reasoning Process - Multi-Hop Question Answering
### Overview
This diagram illustrates a multi-hop question answering process using a Large Language Model (LLM). It depicts the LLM's reasoning steps to answer the question: "How many months passed between Andrew adopting Toby and Buddy?". The process is broken down into stages: Full-Context, Retrieve-then-Answer, and MemR (Memory Retrieval) with two sub-stages (1/2) and (2/2). Each stage includes a query, retrieval actions, memories, evidence, gaps, and an answer. The diagram highlights the LLM's attempts, including incorrect answers and the refinement of queries to achieve the correct result.
### Components/Axes
The diagram is structured into four main sections, arranged horizontally. Each section represents a stage in the reasoning process. Within each section, the following elements are present:
* **Query (Q):** The question or refined query being addressed.
* **Action:** The action the LLM takes (e.g., Retrieve, Reflect).
* **Memories:** Relevant memories retrieved by the LLM.
* **Evidence:** Facts extracted from the memories.
* **Gaps:** Missing information needed to answer the question.
* **Answer:** The LLM's attempt at answering the question, with an indication of correctness (correct/wrong).
* **Visual Indicators:** Red "X" marks incorrect answers, and a green checkmark indicates a correct answer.
### Detailed Analysis or Content Details
**1) Full-Context:**
* **Query:** How many months passed between Andrew adopting Toby and Buddy?
* **Action:** Retrieve [all memories]
* **Memories:** Not explicitly listed, but implied to be all available memories.
* **Evidence:** Andrew adopted Toby on July 11, 2023, and another pup was adopted near October 19, 2023.
* **Gaps:** The name of another pup is unknown, and lacks the specific adoption date for Buddy.
* **Answer:** Six months passed between Andrew adopting Toby and Buddy. [wrong]
* **Note:** A large red "X" is overlaid, indicating an incorrect answer. A text box states "Heavy context reduces LLM's performance".
**2) Retrieve-then-Answer:**
* **Query:** q â Îq
* **Action:** Retrieve [q-relevant memories]
* **Memories:**
1. (11 July, 2023) Andrew: Hey! So much has changed since last time we talked â meet Toby, my puppy.
2. (19 October, 2023) Andrew: Speaking of which, weâve got some awesome news â I recently adopted another pup from a shelter. Heâs the best.
* **Evidence:** Andrew adopted Toby on July 11, 2023, and another pup was adopted on October 19, 2023.
* **Gaps:** It lacks the specific adoption date for Buddy.
* **Answer:** About 4 months. [wrong]
**3) MemR (1/2):**
* **Query:** q â Îq
* **Action:** Retrieve [q-relevant memories]
* **Memories:** [the same as that in 2)]
* **Evidence:** Andrew adopted Toby on July 11, 2023, and another pup was adopted near October 19, 2023.
* **Gaps:** The name of another pup is unknown, and it lacks the specific adoption date for Buddy.
* **Action:** Retrieve
* **Refined Query:** Îq: Buddy adoption date
* **New Query:** q* = q + Îq
* **Action:** Retrieve [q*-relevant memories]
* **Memories:**
1. (19 October, 2023) Andrew: I named him Buddy because heâs my buddy and I hope him and Toby become buddies!
2. âŠ
* **Evidence:** Andrew adopted Toby on July 11, 2023, and Buddy was named on October 19, 2023.
* **Gaps:** It lacks the specific adoption date for Buddy.
**4) MemR (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.
* **Action:** Reflect
* **Updated Evidence:** [The same as above].
* **Updated 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 the 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.
* **Answer:** 3 months. [correct]
* **Note:** A green checkmark indicates a correct answer.
### Key Observations
* The LLM initially struggles with the question, providing incorrect answers in the Full-Context and Retrieve-then-Answer stages.
* The MemR stages demonstrate the LLM's ability to refine its query and retrieve more relevant information.
* The "Heavy context reduces LLM's performance" note suggests that providing too much initial information can hinder the LLM's reasoning.
* The final answer is achieved through a multi-step process of retrieval, reflection, and calculation.
### Interpretation
This diagram illustrates the iterative nature of LLM reasoning for complex questions. The LLM doesn't immediately arrive at the correct answer but refines its approach through multiple steps. The MemR stages highlight the importance of targeted memory retrieval and the ability to synthesize information from different sources. The diagram demonstrates that even with access to relevant information, LLMs may require explicit guidance and refinement of queries to achieve accurate results. The process shows how the LLM breaks down a complex question into smaller, manageable steps, ultimately leading to a correct answer. The diagram also suggests that LLM performance can be affected by the amount of initial context provided, emphasizing the need for efficient information retrieval strategies. The diagram is a visual representation of the LLM's internal thought process, providing insights into its strengths and limitations.
</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 evidenceâgap tracker (Acts 0â3), refines the query about Buddyâs 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 agentâs 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) Evidenceâgap state abstraction for explainable retrieval. MemR 3 maintains a global evidenceâgap 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 agentâs 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 AtkinsonâShiffrin 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-cacheâbased 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 evidenceâgap 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 âLLM-as-controllerâ 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 evidenceâgap 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
\n
## Diagram: LLM Reasoning Process Flow
### Overview
The image depicts a diagram illustrating the process flow of a Large Language Model (LLM) reasoning system. It shows how a user query interacts with a memory module, an LLM generation component, and a router to arrive at a final answer. The diagram highlights the iterative nature of the process, involving retrieval, reflection, and answering steps.
### Components/Axes
The diagram is segmented into four main areas: "Input Data", "LLM Generation", "Final Answer", and "Router". Within these areas are several components:
* **Input Data:** Contains a "User Query" box and a "Memory đ" section, which is further divided into "Chunk-based" (represented by stacked cylinders) and "Graph-based" (represented by a network of nodes and edges).
* **LLM Generation:** Includes an "Evidence-gap Tracker" with "Evidence" and "Gap" sections, and a "Generated Actions" section with three actions: "Retrieve", "Reflect", and "Answer".
* **Final Answer:** Displays the "Answer" and associated steps: "Reflect", "Answer", and "End".
* **Router:** Contains three parameters: "Iteration budget", "Reflect-streak capacity", and "Retrieval opportunity check".
* **Flow Arrows:** Indicate the direction of information flow between components.
* **Text Labels:** Numerous text labels describe the function of each component and the actions performed.
* **Equation:** `q^ret = q â Îq` is present at the bottom of the diagram.
### Detailed Analysis or Content Details
The diagram illustrates the following process:
1. **Input:** A "User Query" is presented: "How many months passed between Andrew adopting Toby and Buddy?".
2. **Memory Retrieval:** The query initiates a "Retrieve" action (green arrow) from the LLM Generation to the "Memory đ". The memory is represented by two structures: "Chunk-based" and "Graph-based".
3. **LLM Generation:** The retrieved information (q^(ret), đ) is fed into the LLM Generation component.
4. **Evidence-Gap Tracking:** The LLM processes the information, tracking "Evidence" ("Andrew adopted Toby on July 11, 2023, and Buddy was named on October 19, 2023.") and identifying a "Gap" ("It lacks the specific adoption date for Buddy.").
5. **Generated Actions:** Based on the evidence and gap, the LLM generates three actions:
* "Retrieve": represented by a magnifying glass icon, suggesting a new query (Îq).
* "Reflect": represented by a brain icon, indicating reasoning (r).
* "Answer": represented by a document icon, indicating a draft answer (w).
6. **Router:** The generated actions are routed based on parameters like "Iteration budget", "Reflect-streak capacity", and "Retrieval opportunity check".
7. **Iteration:** The "Reflect" and "Answer" actions loop back into the LLM Generation, creating an iterative process.
8. **Final Answer:** After sufficient iterations, the process reaches the "Final Answer" stage, providing the answer: "The total number of full months that passed between Andrew adopting Toby and Buddy is 3 months."
9. **End:** The process concludes with an "End" step.
The equation `q^ret = q â Îq` suggests that the retrieved query (q^ret) is a combination of the original query (q) and a delta query (Îq) obtained through the retrieval process.
### Key Observations
* The diagram emphasizes the iterative nature of the LLM reasoning process.
* The "Evidence-gap Tracker" highlights the importance of identifying missing information.
* The "Router" component suggests a mechanism for controlling the iterative process.
* The diagram provides a high-level overview of the process without delving into the specific algorithms or techniques used.
* The diagram is visually clean and uses icons to represent different actions.
### Interpretation
The diagram illustrates a sophisticated LLM reasoning system that goes beyond simple question answering. It demonstrates a process of iterative refinement, where the LLM actively seeks out missing information, reasons about the available evidence, and generates a draft answer. The "Router" component suggests a control mechanism to prevent infinite loops or inefficient iterations. The diagram highlights the importance of both knowledge retrieval and reasoning in achieving accurate and reliable answers. The use of an "Evidence-gap Tracker" is a key feature, indicating a focus on identifying and addressing uncertainties in the available information. The equation `q^ret = q â Îq` formalizes the idea that the retrieval process is not simply about finding relevant information, but about refining the original query based on what is learned. This suggests a dynamic and adaptive approach to knowledge retrieval. The diagram is a conceptual model, and the specific implementation details may vary. However, it provides a valuable framework for understanding the key components and processes involved in LLM reasoning.
</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 evidenceâgap 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 âgapâ) 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
\n
## Screenshot: Query and Response Interface
### Overview
The image is a screenshot of a query and response interface, likely from a conversational AI or information retrieval system. It displays a user query and the system's response, broken down into "Evidence" and "Gaps" sections.
### Components/Axes
The interface is divided into two main sections: "Query" and "Response". Each section is contained within a rectangular box with a light gray background. The "Query" box is positioned above the "Response" box. Within the "Response" box, there are two sub-sections: "Evidence" and "Gaps", each also contained within a rectangular box. Small icons are present to the left of each section title.
### Content Details
**Query:**
The query text is: "What happened 2 days after my last dentist appointment?"
An icon of a person with a speech bubble is present to the left of the query.
**Response:**
* **Evidence:** "You had a dentist appointment on July 12."
An icon of a magnifying glass is present to the left of the evidence.
* **Gaps:** "Information about events on July 14 is missing. Whether July 12 is indeed the most recent dentist appointment is unknown."
An icon of a broken chain link is present to the left of the gaps.
### Key Observations
The system provides a direct answer based on available information (Evidence) and explicitly states what information is missing to fully answer the query (Gaps). The response highlights the uncertainty regarding the recency of the July 12 appointment.
### Interpretation
The interface demonstrates a system capable of understanding a natural language query, retrieving relevant information, and acknowledging limitations in its knowledge. The separation of "Evidence" and "Gaps" is a valuable feature, promoting transparency and helping the user understand the basis for the response. The system doesn't attempt to infer or guess information it doesn't have, instead explicitly stating the missing data. This approach is crucial for building trust and avoiding misinformation. The system is designed to be informative and honest about its capabilities. The use of icons adds a visual cue to the type of information being presented.
</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 agentâs progress in answering the query.
Through the evidence-gap tracker, MemR 3 maintains a structured and transparent internal state that continuously refines the agentâs 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 $\mathcal{G}_{k}\neq\varnothing$ , 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 agentâs 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 $r\in R(q)$ 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 routerâs 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 $k\geq n_{\text{max}}$ 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 $n_{\text{streak}}\geq n_{\text{cap}}$ 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 âadversarialâ 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, âFull-Contextâ 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 evidenceâgap 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 modelâs 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 Zepâs 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 âwhat is already knownâ versus âwhat is still missingâ helps the agent align and compare temporal relations more robustly.
Open-Domain. Open-domain questions are less tied to the userâs 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 routerâs 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 $n_{\text{chk}}\in\{1,3,5,7,9\}$ . In Fig. 4(b), we fix $n_{\text{chk}}=3$ and perform ablations over $n_{\text{max}}\in\{1,2,3,4,5\}$ . 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
\n
## Line Chart: LLM-as-a-Judge Performance vs. Chunk Size
### Overview
This line chart depicts the performance of a Large Language Model (LLM) used as a judge, measured as a percentage, across four different categories of tasks. The performance is plotted against the number of chunks per iteration, ranging from 1 to 9. The chart illustrates how performance changes as the chunk size increases for each category.
### Components/Axes
* **X-axis:** "# chunks / iteration" - Ranges from 1 to 9, with markers at 1, 3, 5, 7, and 9.
* **Y-axis:** "LLM-as-a-Judge (%)" - Ranges from 0 to 100, with markers at 0, 20, 40, 60, 80, and 100.
* **Legend:** Located in the top-right corner, identifying the four categories:
* Multi-hop (Teal dashed line)
* Temporal (Blue square line)
* Open-domain (Peach line)
* Single-hop (Magenta dashed line)
### Detailed Analysis
* **Single-hop (Magenta):** The line starts at approximately 84% at chunk size 1 and increases slightly to around 87% at chunk size 9. The trend is generally flat, with a slight upward slope.
* **Temporal (Blue):** The line begins at approximately 65% at chunk size 1, rises to around 74% at chunk size 3, plateaus around 76% at chunk size 5, and remains relatively stable at approximately 76-77% for chunk sizes 7 and 9. The trend is initially upward, then flattens.
* **Open-domain (Peach):** The line starts at approximately 58% at chunk size 1, increases to around 63% at chunk size 3, then declines to approximately 59% at chunk size 9. The trend is initially upward, then downward.
* **Multi-hop (Teal):** The line begins at approximately 70% at chunk size 1, increases to around 75% at chunk size 3, plateaus around 76% at chunk size 5, and remains relatively stable at approximately 76-77% for chunk sizes 7 and 9. The trend is initially upward, then flattens.
### Key Observations
* Single-hop consistently exhibits the highest performance across all chunk sizes.
* Temporal and Multi-hop show similar performance trends, increasing initially and then plateauing.
* Open-domain demonstrates a different pattern, with an initial increase followed by a decline in performance.
* The performance differences between categories become more pronounced at larger chunk sizes (7 and 9).
### Interpretation
The data suggests that increasing the number of chunks per iteration generally improves the LLM's performance as a judge, particularly for Temporal and Multi-hop tasks. However, beyond a certain point (around 3-5 chunks), the performance gains diminish. The Open-domain category's declining performance with increasing chunk size is an anomaly, potentially indicating that larger chunks introduce noise or complexity that negatively impacts the LLM's ability to judge in this domain. The consistently high performance of Single-hop suggests that this type of task is less sensitive to chunk size or that the LLM is particularly well-suited for it. The chart highlights the importance of optimizing chunk size for specific task categories to maximize LLM performance. The plateauing of Temporal and Multi-hop suggests a point of diminishing returns, where further increasing chunk size does not yield significant improvements.
</details>
(a)
<details>
<summary>x5.png Details</summary>
### Visual Description
\n
## Line Chart: LLM-as-a-Judge Performance vs. Iterations
### Overview
This image presents a line chart illustrating the performance of a Large Language Model (LLM) as a judge, measured as a percentage, across four different categories of tasks (Multi-hop, Temporal, Open-domain, and Single-hop) as the maximum number of iterations increases from 1 to 5.
### Components/Axes
* **X-axis:** "max iterations" - ranging from 1 to 5.
* **Y-axis:** "LLM-as-a-judge (%)" - ranging from 0 to 80, with markings at 20, 40, 60, and 80.
* **Legend:** Located in the center-right of the chart, listing the four categories:
* Multi-hop (dashed blue line)
* Temporal (blue square line)
* Open-domain (orange line with circle markers)
* Single-hop (pink dashed line)
### Detailed Analysis
Let's analyze each data series:
* **Multi-hop (dashed blue line):** The line starts at approximately 72% at iteration 1, increases slightly to around 74% at iteration 2, remains relatively stable around 73-74% for iterations 3 and 4, and then increases to approximately 75% at iteration 5.
* **Temporal (blue square line):** The line begins at approximately 73% at iteration 1, increases to around 75% at iteration 2, decreases slightly to 74% at iteration 3, increases to approximately 76% at iteration 4, and then decreases to around 75% at iteration 5.
* **Open-domain (orange line with circle markers):** The line starts at approximately 58% at iteration 1, increases gradually to around 60% at iteration 2, remains relatively stable around 60-61% for iterations 3 and 4, and then increases to approximately 62% at iteration 5.
* **Single-hop (pink dashed line):** The line is consistently high, starting at approximately 85% at iteration 1 and remaining nearly constant at around 85-86% for all iterations up to 5.
### Key Observations
* The "Single-hop" category consistently exhibits the highest performance, remaining above 85% across all iterations.
* The "Open-domain" category consistently exhibits the lowest performance, remaining below 62% across all iterations.
* The "Multi-hop" and "Temporal" categories show similar performance levels, fluctuating between approximately 73% and 76%.
* All categories show a slight positive trend in performance as the number of iterations increases, although the effect is more pronounced for "Open-domain".
### Interpretation
The data suggests that the LLM-as-a-judge performs best on "Single-hop" tasks and least well on "Open-domain" tasks. The increase in performance with more iterations indicates that allowing the LLM more opportunities to refine its judgment improves accuracy, particularly for more complex tasks like "Open-domain". The relatively stable performance of "Single-hop" suggests that these tasks are easily judged even with a limited number of iterations. The slight improvements observed across all categories with increasing iterations suggest that iterative refinement is a generally beneficial strategy for LLM-based judgment. The difference in performance between the categories likely reflects the inherent difficulty of each task type, with "Open-domain" requiring more nuanced reasoning and knowledge than "Single-hop".
</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 (2â4 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 $r\in R(q)$ is supported by some retrieved memory item $m\in\bigcup_{j\leq k}S_{j}$ 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 â When did Melanie paint a sunrise? â, the correct answer in our setup is simply â 2022 â (the year). MemR 3 quickly finds this year at the first iteration based on evidence â Melanie painted the lake sunrise image last year (2022). â. However, under the idealized abstraction, the requirement set $R(q)$ implicitly includes an exact date predicate (yearâmonthâday), and no memory item $m\in\bigcup_{j\leq K}S_{j}$ 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 questionâs granularity and the available memory, rather than a failure of the agent.
<details>
<summary>x6.png Details</summary>
### Visual Description
\n
## Bar Chart: Iteration Count by Category
### Overview
This is a bar chart visualizing the iteration count for different categories across five iterations. The chart compares the number of iterations performed for "Multi-hop", "Temporal", "Open-Domain", and "Single-hop" categories. The y-axis represents the iteration count, and the x-axis represents the categories. Each category has five bars, one for each iteration.
### Components/Axes
* **X-axis:** Categories - Multi-hop, Temporal, Open-Domain, Single-hop.
* **Y-axis:** Iteration Count - Scale ranges from 0 to 700, with increments of 50.
* **Legend:** Located in the top-right corner, indicating the color-coding for each iteration:
* Iteration 1: Dark Gray
* Iteration 2: Medium Gray
* Iteration 3: Light Gray
* Iteration 4: Light Blue
* Iteration 5: Pale Blue
* **Horizontal Lines:** Two horizontal dashed lines are present at approximately y = 650.
### Detailed Analysis
The chart consists of four groups of bars, one for each category. Within each group, there are five bars representing the iteration counts for iterations 1 through 5.
* **Multi-hop:**
* Iteration 1: Approximately 184
* Iteration 2: Approximately 1
* Iteration 3: Approximately 2
* Iteration 4: Approximately 0
* Iteration 5: Approximately 0
* **Temporal:**
* Iteration 1: Approximately 168
* Iteration 2: Approximately 95
* Iteration 3: Approximately 4
* Iteration 4: Approximately 1
* Iteration 5: Approximately 0
* **Open-Domain:**
* Iteration 1: Approximately 36
* Iteration 2: Approximately 2
* Iteration 3: Approximately 0
* Iteration 4: Approximately 0
* Iteration 5: Approximately 0
* **Single-hop:**
* Iteration 1: Approximately 690
* Iteration 2: Approximately 1
* Iteration 3: Approximately 1
* Iteration 4: Approximately 58
* Iteration 5: Approximately 137
The bars for Iteration 1 are significantly higher than those for subsequent iterations across all categories. The bars for Iterations 4 and 5 are generally low, except for Single-hop.
### Key Observations
* The "Single-hop" category has a dominant number of iterations in the first iteration (approximately 690), far exceeding any other category or iteration.
* Iteration 1 consistently shows the highest iteration counts across all categories.
* Iterations 2, 3, 4, and 5 show significantly lower iteration counts, suggesting a decrease in activity or a convergence of results after the first iteration.
* The "Open-Domain" category consistently has the lowest iteration counts across all iterations.
* The "Temporal" category shows a noticeable decrease in iteration count from Iteration 1 (168) to Iteration 2 (95), and then a sharp decline in subsequent iterations.
### Interpretation
The data suggests that the initial iteration (Iteration 1) requires the most effort or generates the most activity across all categories. This could be due to initial setup, data exploration, or the identification of initial issues. The subsequent decrease in iteration counts indicates that the process converges or stabilizes after the first iteration.
The "Single-hop" category's high iteration count in the first iteration, and its subsequent increase in Iterations 4 and 5, might indicate that this category is more complex or requires more refinement than the others. The "Open-Domain" category consistently having low iteration counts suggests it may be a simpler or less problematic area.
The horizontal lines at y=650 may represent a threshold or target value. The "Single-hop" category in Iteration 1 exceeds this threshold significantly, while other categories fall well below it. This could indicate that the "Single-hop" category is performing exceptionally well or requires further attention.
</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 datasetâs 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 âdecision_directiveâ instructs the maximum iteration budges, reflect-streak capacity, and retrieval opportunity check, introduced in Sec. 3.4. Generally, âdecision_directiveâ is a textual instruction: âreflectâ if you need to think about the evidence and gaps; choose âanswerâ ONLY when evidence is solid and no gaps are noted; choose âretrieveâ otherwise. However, when the maximum iterations budget is reached, âdecision_directiveâ is set as âanswerâ to stop early. When the reflection reaches the maximum capacity, âdecision_directiveâ is set as âretrieveâ to avoid repeated ineffective reflection. When there is no useful retrieval remains, âdecision_directiveâ is set as âreflectâ 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 userâs 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 â âgapsâ will do that. - gaps: gaps between the question and evidence that prevent a complete answer. - decision: one of [âretrieveâ,âanswerâ,âreflectâ]. Choose {decision_directive}. Only include these conditional keys: - retrieval_query: only when decision == âretrieveâ. Provide a STANDALONE search string; short (5-15 tokens). * BAD Query: âthe dateâ (lacks context). * GOOD Query: âgraduation ceremony dateâ (specific). * STRATEGY: 1. Search for the ANCHOR EVENT. (e.g. Question: âWhat happened 2 days after X?â, Query: âtimestamp of event Xâ). 2. Search for the MAPPED ENTITY. (e.g. Question: âWeather in the Windy Cityâ, Query: âweather in Chicagoâ). - detailed_answer: only when decision == âanswerâ; 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 == âreflectâ; 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, âquestionâ is the original question $q$ . âevidence_blockâ and âgap_blockâ are evidence $\mathcal{E}_{k}$ and gaps $\mathcal{G}_{k}$ introduced in Sec. 3.3. âraw_blockâ is the retrieved raw snippets $\mathcal{S}_{k}$ in Eq. 5. âreasoning_blockâ is the reasoning content $\mathcal{F}_{k}$ in Sec. 3.4. âlast_queryâ 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 âNoneâ 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 âHow many months passed between events $A$ and $B$ ?â, the requirement set can be
$$
R(q)=\{\text{date}(A),\text{date}(B)\}. \tag{8}
$$
Each requirement $r\in R(q)$ 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 $S_{k}\subseteq\mathcal{M}$ denote the snippets retrieved at iteration $k$ . We define a relation $m\models r$ to indicate that memory item $m\in\mathcal{M}$ contains sufficient information to support requirement $r\in R(q)$ . 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 $E_{k}\subseteq R(q)$ 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 $r\in R(q)$ , we have $r\in E_{k}^{\star}$ if and only if there exists some $m\in\bigcup_{j\leq k}S_{j}$ such that $m\models r$ . Then the following hold:
1. Monotonicity: $E_{k-1}^{\star}\subseteq E_{k}^{\star}$ and $G_{k}^{\star}\subseteq G_{k-1}^{\star}$ for all $k\geq 1$ .
1. Soundness: If $m\models r$ for some retrieved memory $m\in S_{k}$ , then $r\in E_{k}^{\star}$ .
1. Completeness at convergence: If every requirement $r\in R(q)$ is supported by some $m\in\bigcup_{j\leq K}S_{j}$ 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 $E_{k-1}^{\star}\subseteq E_{k}^{\star}$ . Since $G_{k}^{\star}=R(q)\setminus E_{k}^{\star}$ and $E_{k-1}^{\star}\subseteq E_{k}^{\star}$ , we obtain $G_{k}^{\star}\subseteq G_{k-1}^{\star}$ . (2) If $m\models r$ for some $m\in S_{k}$ , then by Definition B.3 we have $r\in\{r^{\prime}\in R(q)\mid\exists m^{\prime}\in S_{k},\;m^{\prime}\models r^{\prime}\}\subseteq E_{k}^{\star}$ . (3) If every $r\in R(q)$ is supported by some $m\in\bigcup_{j\leq K}S_{j}$ 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 $G_{K}^{\star}=R(q)\setminus E_{K}^{\star}=\varnothing$ . â
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 requirementâsupport 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 $r\in R(q)$ 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 â Melanie painted a sunrise â case in Sec. 4.3 provides a concrete illustration of granularity bias. The question asks â When did Melanie paint a sunrise? â, 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 yearâmonthâday of the painting event. The memory store only contains a coarse statement such as â Melanie painted the lake sunrise image last year (2022). â
In the ideal tracker, no memory item $m$ satisfies $m\models r_{\text{date}}$ , so the precondition of Theorem B.4 âs 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 âfailuresâ 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, âFull-Contextâ 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 LangChainâs 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 modelâs 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 âs 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 Joleneâs snakes? (conversation 8, question 17), (conversation 8, question 90)
1. What are Joleneâs 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 Deborahâs 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 Joleneâs 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 $\pm$ std | 69.39 $\pm$ 0.41 | 73.83 $\pm$ 1.40 | 67.01 $\pm$ 1.64 | 80.60 $\pm$ 0.18 | 76.26 $\pm$ 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 $\pm$ std | 71.39 $\pm$ 1.08 | 76.22 $\pm$ 0.95 | 61.11 $\pm$ 1.59 | 89.44 $\pm$ 0.18 | 81.56 $\pm$ 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 $\pm$ std | 77.78 $\pm$ 1.44 | 77.78 $\pm$ 0.26 | 69.79 $\pm$ 1.04 | 84.42 $\pm$ 0.12 | 80.88 $\pm$ 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 $\pm$ std | 81.20 $\pm$ 1.62 | 82.14 $\pm$ 1.29 | 71.53 $\pm$ 3.01 | 92.17 $\pm$ 0.42 | 86.75 $\pm$ 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
\n
## Chart: Retrieved Token Usage vs. LLM-as-a-Judge Score
### Overview
The image presents a combined bar and line chart comparing the retrieved token usage and LLM-as-a-judge score for three different methods: RAG, MemR3, and Full-Context. The chart visualizes how token usage varies across different categories (Multi-hop, Temporal, Open-domain, Single-hop) and how these relate to the LLM-as-a-judge score. The Y-axis on the left represents Retrieved Token Usage on a logarithmic scale, while the Y-axis on the right represents the LLM-as-a-Judge Score in percentage.
### Components/Axes
* **X-axis:** Methods - RAG, MemR3, Full-Context.
* **Left Y-axis:** Retrieved Token Usage (Logarithmic Scale). Ranges from 10â° (1) to 10â” (100,000).
* **Right Y-axis:** LLM-as-a-Judge Score (%). Ranges from 50% to 100%.
* **Legend (Top-Right):**
* Multi-hop (Teal Circle)
* Temporal (Blue Square)
* Open-domain (Orange Star)
* Single-hop (Pink Diamond)
* LLM-as-a-Judge (Black Line)
* **Bars:** Represent Retrieved Token Usage for each category and method. The bars are colored according to the categories.
* **Lines:** Represent LLM-as-a-Judge Score for each method.
### Detailed Analysis
The chart displays the following data:
**RAG (Retrieval-Augmented Generation):**
* Multi-hop: Retrieved Token Usage â 75, LLM-as-a-Judge Score â 72%
* Temporal: Retrieved Token Usage â 250, LLM-as-a-Judge Score â 78%
* Open-domain: Retrieved Token Usage â 10, LLM-as-a-Judge Score â 60%
* Single-hop: Retrieved Token Usage â 1000, LLM-as-a-Judge Score â 88%
* LLM-as-a-Judge (overall): â 75%
**MemR3:**
* Multi-hop: Retrieved Token Usage â 150, LLM-as-a-Judge Score â 82%
* Temporal: Retrieved Token Usage â 800, LLM-as-a-Judge Score â 85%
* Open-domain: Retrieved Token Usage â 10, LLM-as-a-Judge Score â 62%
* Single-hop: Retrieved Token Usage â 1200, LLM-as-a-Judge Score â 92%
* LLM-as-a-Judge (overall): â 85%
**Full-Context:**
* Multi-hop: Retrieved Token Usage â 120, LLM-as-a-Judge Score â 75%
* Temporal: Retrieved Token Usage â 500, LLM-as-a-Judge Score â 80%
* Open-domain: Retrieved Token Usage â 8, LLM-as-a-Judge Score â 58%
* Single-hop: Retrieved Token Usage â 1000, LLM-as-a-Judge Score â 90%
* LLM-as-a-Judge (overall): â 82%
**Trends:**
* **LLM-as-a-Judge Line:** The black line representing the LLM-as-a-Judge score generally slopes upward from RAG to MemR3, then slightly decreases towards Full-Context.
* **Single-hop:** Consistently shows the highest token usage across all methods.
* **Open-domain:** Consistently shows the lowest token usage across all methods.
### Key Observations
* MemR3 consistently achieves the highest LLM-as-a-Judge scores.
* Single-hop queries require the most tokens across all methods.
* Open-domain queries require the fewest tokens.
* There's a general trend of increasing LLM-as-a-Judge score with increasing token usage, but this isn't strictly linear. The relationship appears to plateau or even decrease slightly for Full-Context.
### Interpretation
The chart suggests that MemR3 is the most effective method for balancing token usage and LLM-as-a-Judge score. While higher token usage doesn't always guarantee a higher score, it generally correlates positively, especially when comparing RAG and MemR3. The Full-Context method shows a slight decrease in LLM-as-a-Judge score despite comparable token usage to MemR3, indicating that simply providing more context isn't always beneficial.
The differences in token usage across categories highlight the varying complexity of different query types. Single-hop queries, likely requiring less information retrieval, use the most tokens, while Open-domain queries, potentially benefiting from broader context, use the least. This could be due to the retrieval mechanisms prioritizing more information for simpler queries or the nature of the data itself.
The logarithmic scale on the Y-axis for token usage is crucial. It emphasizes the significant differences in token usage between categories like Open-domain and Single-hop. The LLM-as-a-Judge score provides a metric for evaluating the quality of the retrieved information, suggesting that the methods are not only retrieving tokens but also retrieving *relevant* tokens. The plateauing effect observed with Full-Context suggests a point of diminishing returns â adding more context beyond a certain threshold doesn't necessarily improve the LLM's judgment.
</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.