Source-linked AI summary
MemMA: Coordinating the Memory Cycle through Multi-Agent Reasoning and In-Situ Self-Evolution
Minhua Lin, Zhiwei Zhang, Hanqing Lu, Hui Liu, Xianfeng Tang, Qi He, Xiang Zhang, Suhang Wang
TL;DR
Long-horizon memory systems often treat construction, retrieval, and utilization as isolated operations, leaving strategic coordination and timely memory repair unresolved. MemMA addresses both gaps with multi-agent guidance and in-situ self-evolution, and experiments on LoCoMo show improvements across backbones and storage backends.
Problem
Existing memory-augmented agents often isolate construction, retrieval, and utilization, creating strategic blindness on the forward path and sparse, delayed supervision on the backward path.
Method
MemMA coordinates the memory cycle with a Meta-Thinker guiding construction and iterative retrieval, plus probe-QA-based verification and repair before memory commitment.
Results
MemMA outperforms existing baselines across multiple LLM backbones and consistently improves three storage backends on LoCoMo.
Takeaways & Limitations
The results support coordinating forward memory execution with backward utilization feedback rather than optimizing memory stages independently.
Takeaways & Limitations
The backward-path design assumes interaction streams can be organized into sessions and synthetic probe QA can provide useful localized supervision.
Abstract
from arXiv · showhide
Memory-augmented LLM agents maintain external memory banks to support long-horizon interaction, yet most existing systems treat construction, retrieval, and utilization as isolated subroutines. This creates two coupled challenges: strategic blindness on the forward path of the memory cycle, where construction and retrieval are driven by local heuristics rather than explicit strategic reasoning, and sparse, delayed supervision on the backward path, where downstream failures rarely translate into direct repairs of the memory bank. To address these challenges, we propose MemMA, a plug-and-play multi-agent framework that coordinates the memory cycle along both the forward and backward paths. On the forward path, a Meta-Thinker produces structured guidance that steers a Memory Manager during construction and directs a Query Reasoner during iterative retrieval. On the backward path, MemMA introduces in-situ self-evolving memory construction, which synthesizes probe QA pairs, verifies the current memory, and converts failures into repair actions before the memory is finalized. Extensive experiments on LoCoMo show that MemMA consistently outperforms existing baselines across multiple LLM backbones and improves three different storage backends in a plug-and-play manner. Our code is publicly available at https://github.com/ventr1c/memma.
1 Introduction
Long-horizon agents need external memory because ephemeral context is costly and attention-diluting, but existing systems often optimize memory stages separately. MemMA coordinates strategic reasoning across construction and retrieval while converting downstream failures into immediate memory repairs.
- Motivation: Persistent agents require actively managed external memory banks to retain and retrieve information over long horizons.Ephemeral context windows are insufficient because they are computationally expensive and prone to attention dilution.
- Challenges: The memory cycle couples construction, retrieval, and utilization, making isolated optimization fundamentally suboptimal.Construction determines stored information, retrieval surfaces evidence, and utilization reveals whether that evidence supports downstream reasoning.
- Challenges: Strategic blindness causes myopic construction and aimless retrieval when agents lack explicit meta-cognition to coordinate actions toward downstream question answering.Myopic construction accumulates or overwrites conflicting facts, while aimless retrieval performs shallow or repetitive searches.
- Challenges: Sparse, delayed utilization feedback makes it difficult to assign downstream failures to earlier memory-writing decisions, allowing omissions and conflicts to persist.Existing reflection and experiential-learning methods rarely convert downstream failures into direct memory-bank repairs.
- Approach: MEMMA uses a Meta-Thinker to guide a Memory Manager and Query Reasoner, while in-situ self-evolution verifies memory with probe QA and repairs failures before commitment.The framework coordinates both forward execution and backward repair through a plug-and-play multi-agent architecture.
- Experiments: MEMMA outperforms existing baselines on LoCoMo across multiple LLM backbones and consistently improves three storage backends as a plug-and-play module.The reported experiments evaluate both broad baseline performance and backend portability.
2 Related Work
Prior memory-augmented agents improve individual stages such as architecture, organization, consolidation, and retrieval, but primarily optimize storage or retrieval components separately.
- Memory-Augmented LLM Agents: Existing work advances memory architecture, organization, consolidation, and retrieval for long-horizon LLM agents.These approaches substantially improve individual stages of the memory pipeline.
3 Preliminaries and Motivation
Long-horizon conversational memory is modeled as a coupled cycle in which construction, retrieval, and utilization jointly determine answer accuracy. Preliminary analysis shows that iterative refinement helps, while explicit strategic guidance provides additional control, motivating coordinated memory operations.
- Task Setup: The task processes dialogue chunks organized into sessions while updating a structured external memory bank and answering questions from retrieved evidence.The agent is evaluated after processing the full stream, with the goal of jointly improving construction and retrieval.
- Memory Cycle Effect as a Design Lens: The memory cycle comprises coupled construction, retrieval, and utilization phases rather than isolated pipeline modules.Construction constrains retrieval, retrieval constrains utilization, and utilization outcomes expose upstream deficiencies.
- Motivating Analysis: Strategic Blindness: Active agents remain strategically blind when local context or similarity signals trigger operations without an explicit global strategy.This limitation affects both memory editing and retrieval-query formulation.
- Motivating Analysis: Strategic Blindness: 54.6% Acc for Unguided Active exceeds 52.6% for Static, indicating that refinement can improve evidence retrieval when initial queries are incomplete or mismatched.These preliminary results use LoCoMo with GPT-4o-mini as the backbone LLM.
- Motivating Analysis: Strategic Blindness: 59.2% Acc for Strategic Active exceeds Unguided Active, supporting the value of explicit strategic guidance when active operators are otherwise unchanged.The analysis connects this gap to both aimless retrieval and myopic construction.
4 Methodology
MEMMA coordinates memory construction, retrieval, and utilization through specialized agents on both forward and backward paths. It separates strategic reasoning from execution and uses in-situ probes to verify and repair memory before commitment.
- Forward-path coordination: A planner–worker architecture separates strategic reasoning from low-level execution across memory construction, retrieval, and answer generation.The Meta-Thinker supplies guidance, while specialized agents perform memory editing, evidence retrieval, and answering.
- Forward-path coordination: During construction, the Meta-Thinker analyzes new dialogue against bounded existing memory and guides atomic edits that retain, consolidate, or resolve information.The Memory Manager selects an action based on the current chunk, bounded memory view, and construction guidance.
- Forward-path coordination: During retrieval, the Query Reasoner iteratively refines searches when evidence is insufficient, targeting diagnosed information gaps until evidence is answerable or the budget is reached.The Meta-Thinker evaluates coverage, consistency, and specificity, while successive queries address missing attributes or temporal scope rather than repeating near-duplicate searches.
- Backward-path self-evolution: In-situ self-evolving construction synthesizes probe QA pairs grounded in the current session and historical context to provide immediate supervision after provisional memory construction.Probes test factual recall, cross-session relational reasoning, and temporal inference.
- Backward-path self-evolution: Failed probes are converted into repair proposals that diagnose missing or hard-to-retrieve information, then consolidated as SKIP, MERGE, or INSERT actions before write-back.Consolidation resolves redundancy and conflicts against existing memory and across repair proposals, keeping the evolving memory compact and internally consistent.
5 Experiments
Experiments on LoCoMo evaluate MEMMA against baselines, across storage backends, and through component and budget analyses. MEMMA improves baseline performance, transfers across backends, and benefits from coordinated construction, retrieval, and repair.
- Main Comparison with Baselines: MEMMALM achieves the best overall performance under both backbones, improving over LightMem by +4.82 F1, +1.62 B1, and +5.92 ACC with GPT-4o-mini.It reaches 49.40 F1, 38.28 B1, and 81.58 ACC under GPT-4o-mini.
- Main Comparison with Baselines: Under Claude-Haiku-4.5, MEMMALM achieves the best overall ACC, improving from 73.03 to 76.97 over LightMem.
- Flexibility across Storage Backends: MEMMA improves ACC across Single-Agent, A-Mem, and LightMem backends, rising from 52.60 to 84.87, 52.63 to 78.29, and 75.66 to 81.58, respectively.The weaker Single-Agent backend shows higher semantic correctness despite a decrease in B1.
- In-depth Dissection of MemMA: Iterative retrieval is the most critical forward-path component, with MEMMASA/R reducing GPT-4o-mini ACC from 84.87 to 70.39.The same ablation reduces Claude-Haiku-4.5 ACC from 88.82 to 81.58.
- In-depth Dissection of MemMA: Self-evolution repairs construction omissions, with MEMMASA/E reducing GPT-4o-mini ACC from 84.87 to 73.68.The large ACC drop with only moderate F1 change indicates a primary effect on semantic correctness.
- In-depth Dissection of MemMA: Retrieval budget effects depend on storage quality: LightMem peaks at k=30–40 with 81.58 ACC, while Single-Agent rises to 84.21 at k=50.Additional retrieval can introduce noise for stronger storage, whereas weaker storage benefits from retrieving more evidence.
- In-depth Dissection of MemMA: Diagnosis-guided refinement converges quickly, as one or two additional retrieval rounds suffice to close most information gaps while further iterations risk retrieval drift.
- In-depth Dissection of MemMA: Case studies show that construction guidance preserves answer-bearing details, while iterative retrieval surfaces missing evidence before answer generation.Iterative retrieval cannot compensate for details that were never preserved during construction.
6 Conclusion
MEMMA coordinates the memory cycle across forward reasoning and backward repair. On LoCoMo, it outperforms baselines across backbones and improves three storage backends.
- MEMMA coordinates the memory cycle along forward and backward paths through a plug-and-play multi-agent framework.
- A Meta-Thinker separates strategic reasoning from low-level execution in construction and retrieval, while in-situ self-evolution repairs memory from probe QA failures before commitment.
- Experiments on LoCoMo show that MEMMA outperforms all baselines across multiple backbones and consistently improves three different storage backends.
7 Limitations
The evaluation is limited to a dialogue-centric long-horizon memory benchmark and assumes session-structured interactions with useful synthetic probe QA supervision.
- LoCoMo covers single-hop, multi-hop, temporal, and open-domain questions but does not represent every setting requiring persistent memory.
- The backward path assumes interaction streams can be organized into sessions and that synthetic probe QA provides useful localized supervision.These assumptions may require adaptation when session boundaries are unclear or interactions are more open-ended.
8 Ethics Statement
The paper studies long-horizon memory management using synthetic conversations rather than real user data and situates its contribution against prior memory and self-improvement systems.
- All experiments use a publicly available benchmark containing synthetic conversations rather than real user data.The paper states that no personally identifiable information is collected, stored, or processed.
- Real-world deployment raises concerns about privacy, informed consent, controllability over stored memories, and persisting incorrect information through automated repair.These concerns are outside the scope of the present study but are identified as design requirements for production deployment.
- Prior work improves memory architecture, organization, retrieval, outputs, experience stores, or memory-use policies, whereas MEMMA directly coordinates the memory cycle and repairs the memory bank.The framework performs construction-level repair through probe QA, verification, evidence-grounded critique, and semantic consolidation without gradient-based training or separate experience stores.
B.1 Evaluation Details
The evaluation details compare progressively more strategic memory pipelines and illustrate why explicit diagnosis and tighter coordination matter for retrieval and construction.
- Baselines: Static performs sequential memory edits and one-shot top-30 cosine retrieval, while Unguided Active adds iterative query rewriting without missing-information diagnosis.Both baselines omit the explicit strategic guidance used by Strategic Active.
- Baselines: Strategic Active adds planner guidance for retaining, consolidating, or resolving memories and for diagnosing missing retrieval evidence.Its active retrieval permits up to five query-rewriting iterations under the baseline setup.
- Retrieval cases: Strategic Active identifies a missing date and retrieves the correct museum memory, whereas unguided rewrites remain lexical paraphrases and drift toward unrelated memories.The example supports diagnosing the information gap rather than merely increasing rewrite rounds.
- Retrieval cases: For an ambiguous conference question, Strategic Active distinguishes a future event from a past one and separates transgender from LGBTQ conferences before one guided rewrite surfaces the relevant memory.The case frames the bottleneck as disambiguation rather than broader search.
- Construction cases: Strategic Active still fragments one support-group episode across overlapping entries because planner guidance is advisory and the Memory Manager edits one utterance at a time.The case motivates global reorganization and tighter coordination beyond local planning.
- In-situ self-evolution: The backward component synthesizes single-hop, multi-session, and temporal probes to expose missing facts, weak cross-session links, and temporal inconsistencies before repair.Failed probes are diagnosed as missing information or difficult retrieval and can lead to candidate repair facts.
F.1 Dataset Details
Experiments use a long-term conversational-memory benchmark with diverse reasoning questions and compare MEMMA against passive retrieval, memory SDK, and structured-memory baselines.
- Dataset: LoCoMo contains 10 conversations spanning roughly 600 dialogue turns and 16K tokens on average, with up to 32 sessions.The full benchmark includes 272 sessions, 5,882 dialogue turns, and 1,986 QA pairs.
- Dataset: The evaluation subset uses one conversation with 19 sessions and 419 dialogue turns, leaving 152 non-adversarial QA pairs across four categories.The categories are single-hop (70), multi-hop (32), temporal (37), and open-domain (13).
- Baselines: The comparison includes Full Text, Naive RAG, LangMem, A-Mem, and LightMem as alternative context, retrieval, and memory-organization approaches.These baselines range from direct context concatenation and fixed-chunk retrieval to structured and multi-stage memory systems.
- Implementation: GPT-4o-mini and Claude-Haiku-4.5 serve as default backbones for the Memory Manager, Meta-Thinker, and Query Reasoner, with GPT-4o-mini fixed for answer generation and judging.The setup uses H=3 query-refinement iterations and generates J=5 probe QA pairs per session with Claude-Opus-4.5.
G.1 Empirical Analysis.
Probe quality strongly affects in-situ memory repair: Opus achieves the best overall repair metrics, while Sonnet’s shorter, mostly single-hop probes provide weaker supervision than Haiku’s and Opus’s more reasoning-intensive probes.
- Probe quality and repair performance: 76.97 ACC and 45.10 F1 make Opus the strongest probe generator overall, outperforming Haiku and Sonnet on repair quality.Haiku and Sonnet both reach 74.34 ACC, while their F1 scores are 44.98 and 43.30, respectively.
- Probe quality and repair performance: Despite identical ACC, Haiku exceeds Sonnet on F1 and B1, indicating higher-quality repairs at the token level.Haiku scores 44.98 F1 and 35.69 B1, compared with Sonnet’s 43.30 F1 and 32.74 B1.
- Probe characteristics: Sonnet produces shorter, more extractive probes, whereas Haiku and Opus generate longer probes with more temporal, multi-session, or cross-session relational questions.Average answer lengths are 11.12 words for Sonnet, 19.43 for Haiku, and 21.48 for Opus.
- Probe characteristics: Sonnet generates more short answers and single-hop questions, which primarily test keyword presence or individual fact storage rather than cross-session consolidation.Sonnet produces 64 of 95 single-hop questions, compared with 25 for Haiku and 26 for Opus.
- Probe characteristics: Haiku and Opus probes more often require multi-attribute reasoning, making them more likely to expose consolidation gaps and trigger deeper repairs.The paper attributes their stronger repair quality to the greater use of multi-hop and cross-session reasoning probes.
H.1 Forward Path: Construction-Time Meta-Thinker Guidance
Forward-path coordination improves memory construction and retrieval by preserving answer-bearing details, preventing harmful fact merges, and refining queries around explicit information gaps before answering.
- Construction-time guidance: Construction-time Meta-Thinker guidance preserves exact answer-bearing details that unguided construction may omit.For Caroline’s neighborhood question, guidance stores the rainbow sidewalk and supporting attributes, while the ablated variant produces a vague, confused answer.
- Construction-time guidance: Meta-Thinker guidance keeps semantically adjacent facts disentangled instead of merging conflicting entries.With guidance, clarinet and violin facts remain distinct; without it, they are incorrectly merged and one fact is overwritten.
- Construction-time guidance: The resulting memory improvements preserve concrete details, avoid destructive merges, and reduce retrieval drift or contradictions before retrieval begins.Additional cases extend this pattern to quoted textual details and topic disentanglement.
- Iterative retrieval: Meta-Thinker-guided retrieval first classifies evidence as ANSWERABLE or NOT-ANSWERABLE, after which the Query Reasoner refines queries to retrieve missing information.The process operates over a fixed memory bank and targets the diagnosed information gap.
- In-situ self-evolution: Self-evolution converts probe failures into repair facts that transfer to downstream benchmark answers through entity insertion, event-detail sharpening, and evidence-cluster completion.These repairs change downstream answers from missing or generic responses to specific entities, events, and objects.