MemMA: Memory-Augmented Multi-Agent Framework

• MemMA is a framework that integrates memory construction, retrieval, and repair to enable coherent long-horizon memory for large language models.
• It uses a multi-agent architecture with specialized roles (Meta-Thinker, Memory Manager, Query Reasoner, Answer Agent) to coordinate the complete memory cycle.
• Experimental evaluations on benchmarks like LoCoMo demonstrate significant gains in metrics such as F1, BLEU, and ACC across diverse storage backends.

Searching arXiv for the specified paper and closely related memory-agent work. MemMA is a plug-and-play multi-agent framework for memory-augmented LLM agents that coordinates the entire memory cycle along both its forward path—construction and retrieval—and its backward path—verification and repair (Lin et al., 19 Mar 2026). It is motivated by the observation that long-horizon agents maintain an external memory bank $M$ to sustain coherent interaction over many sessions, yet most existing systems treat memory construction, retrieval, and utilization as isolated subroutines. In the MemMA formulation, this isolation produces two coupled obstacles: strategic blindness on the forward path, 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 repair of the memory bank. MemMA addresses these obstacles through multi-agent planning and execution, diagnosis-guided iterative retrieval, and in-situ self-evolving memory construction that converts probe failures into localized repair actions before memory is finalized (Lin et al., 19 Mar 2026).

1. Problem formulation and the memory cycle effect

MemMA is designed for long-horizon LLM agents whose memory bank $M$ stores structured entries containing text and metadata such as timestamp, source, and speaker, while a stream of chunks $C$ is organized into sessions $S$ and later queried through retrieved evidence from $M$ (Lin et al., 19 Mar 2026). Its central premise is that memory construction, retrieval, and utilization are tightly coupled: poor construction constrains retrieval, retrieval outcomes constrain utilization, and utilization failures should feed back to improve construction. The paper describes this closed-loop perspective as the “memory cycle effect” (Lin et al., 19 Mar 2026).

The forward-path failure mode is termed strategic blindness. In construction, this appears as “Myopic Construction,” where redundant or conflicting facts are accumulated without resolution. In retrieval, it appears as “Aimless Retrieval,” where shallow, repetitive query rewrites drift without closing the true information gap. The backward-path failure mode is sparse, delayed supervision: wrong answers seldom yield direct credit assignment to earlier construction decisions, so flawed memories propagate (Lin et al., 19 Mar 2026).

This formulation places MemMA within a broader line of memory-agent research, but with a specific emphasis on end-to-end coordination rather than isolated optimization of storage, organization, or retrieval. The paper explicitly contrasts prior systems that optimize storage, organization, or retrieval largely in isolation with MemMA’s attempt to coordinate the entire memory cycle (Lin et al., 19 Mar 2026). A plausible implication is that MemMA treats memory quality as an emergent property of the whole loop rather than of any single module.

2. Multi-agent architecture and formalization

MemMA decomposes memory management into four agent roles: a Meta-Thinker $\pi_p$, a Memory Manager $\pi_s$, a Query Reasoner $\pi_r$, and an Answer Agent $\pi_a$ (Lin et al., 19 Mar 2026). The Meta-Thinker acts as planner. During construction it takes the current chunk $c_t$ and a bounded memory view $M$0 and produces construction guidance $M$1 highlighting importance, redundancy, and conflicts. During retrieval it takes the query $M$2, current evidence $M$3, query history $M$4, and a bounded view $M$5, and produces retrieval guidance $M$6 that judges ANSWERABLE versus NOT-ANSWERABLE, diagnoses missing evidence slots, and specifies constraints or avoid terms for the next query (Lin et al., 19 Mar 2026).

The construction-side worker is the Memory Manager. It consumes $M$7 and chooses an atomic edit

$M$8

with state transition

$M$9

The retrieval-side worker is the Query Reasoner, which generates the next query rewrite targeted to the diagnosed top gap:

$C$0

The Answer Agent receives the final evidence $C$1 and outputs

$C$2

It is kept frozen to decouple answer-generation capacity from memory quality; GPT-4o-mini is used as both answer agent and judge in the reported experiments (Lin et al., 19 Mar 2026).

Retrieval itself is backend-specific but abstracted by a common search interface. A simple formalization given in the paper is

$C$3

where top-$C$4 entries are returned according to the backend’s retrieval scoring, such as cosine similarity over embeddings for vector backends (Lin et al., 19 Mar 2026). This abstraction is what enables MemMA’s plug-and-play integration across different memory stores.

3. Forward-path coordination: construction, retrieval, and answerability diagnosis

On the forward path, MemMA coordinates both memory construction and iterative retrieval. During construction, the Meta-Thinker inspects each chunk $C$5 against a bounded memory view and produces guidance that highlights important factual units to preserve, including names, dates, and objects, while also identifying redundancy and potential conflicts. The Memory Manager then applies backend-specific edits under that guidance (Lin et al., 19 Mar 2026). The explicit purpose is to preserve answer-bearing details and avoid destructive merges.

During retrieval, the process begins with $C$6 and $C$7. The Meta-Thinker then performs answerability diagnosis on $C$8 with bounded $C$9. If the evidence is sufficient, it returns ANSWERABLE; otherwise it returns NOT-ANSWERABLE and emits guidance specifying the top gap, constraints, and avoid terms. The Query Reasoner then generates one orthogonal rewrite targeting that gap, retrieves again, and the loop repeats until the Meta-Thinker returns ANSWERABLE or a refinement budget $S$0 is reached, with $S$1 typically 1–3 (Lin et al., 19 Mar 2026).

The answerability check uses a blocking-gap test: NOT-ANSWERABLE is returned only when a missing fact or contradiction would change the final short answer; otherwise the system should return ANSWERABLE at best-supported granularity (Lin et al., 19 Mar 2026). This criterion is supplemented by explicit decision rules. The anti-stall rule states that if there are at least three queries and the gap is non-blocking, the system should prefer ANSWERABLE at best granularity. For time questions, exact day is required only when asked; otherwise unambiguous granularity such as month or year is acceptable. For “who/what” questions, one clearly supported entity is enough unless exhaustive enumeration is requested (Lin et al., 19 Mar 2026).

The case studies reported in the paper clarify the role of this coordination. In a museum-date example, unguided rewrites such as “visit,” “trip,” and “history” drift, whereas the Meta-Thinker diagnoses a missing date and directs the Query Reasoner toward the exact evidence. In a case involving “transgender” versus “LGBTQ conference,” guidance disambiguates event names and temporal scope. In a “rainbow sidewalk” example, construction guidance preserves an object-level detail that unguided memory would reduce to vague “cool stuff.” In an instruments example, guidance prevents destructive overwriting between “clarinet” and “violin” by preserving parallel facts (Lin et al., 19 Mar 2026).

4. Backward-path self-evolution and memory repair

MemMA’s backward path introduces in-situ self-evolving memory construction. After a session $S$2, an initial construction pass produces provisional memory $S$3, after which the system synthesizes a probe set

$S$4

grounded in the current session and relevant historical memory (Lin et al., 19 Mar 2026). The reported probes cover Single-hop Factoid, Multi-session Reasoning, and Temporal Reasoning, targeting failure modes such as missing entities, incomplete details, weak cross-session links, and temporal inconsistency.

Verification is then performed by retrieving top-$S$5 evidence for each probe and answering with the frozen answer agent:

$S$6

Failures are identified when $S$7 is judged incorrect relative to $S$8 by an LLM judge (Lin et al., 19 Mar 2026). For failed probes, reflection produces candidate repairs $S$9 conditioned on $M$0, explicitly diagnosing “missing info” versus “hard-to-retrieve as written,” and proposing a concrete, retrieval-friendly fact with names, dates, distinctive details, and a target speaker (Lin et al., 19 Mar 2026).

Repairs are consolidated through discrete actions. Each $M$1 is checked against $M$2 and assigned SKIP, MERGE, or INSERT. SKIP marks redundancy; MERGE adds a missing detail to an existing entry about the same event or attribute; INSERT is used for a novel entry, with an explicit rule that different event or time implies INSERT and never MERGE (Lin et al., 19 Mar 2026). Finalization is

$M$3

The paper also gives the conceptual objective

$M$4

implemented non-parametrically through prompting and discrete actions rather than gradient optimization (Lin et al., 19 Mar 2026).

The examples emphasize that this backward loop repairs memory before errors propagate. A probe failure about the artist at Melanie’s daughter’s birthday concert inserts “Matt Patterson,” enabling later answers such as “Summer Sounds” and “Matt Patterson.” A camping-trip probe repairs “Perseid meteor shower,” replacing a generic description with a precise detail. Pottery probes add “colorful bowl” and “black-and-white bowl,” converting vague downstream retrieval into complete lists (Lin et al., 19 Mar 2026). This suggests that MemMA’s self-evolution mechanism improves both semantic correctness and retrieval friendliness.

5. Storage backends, prompting, and deployment model

MemMA is explicitly plug-and-play across three storage backends: Single-Agent, A-Mem, and LightMem (Lin et al., 19 Mar 2026). In the Single-Agent backend, memory is a simple vector-store with top-$M$5 cosine retrieval and straightforward ADD/UPDATE/DELETE operations. In A-Mem, the backend uses a note-based Zettelkasten organization with activation-based retrieval, so atomic edits map to note updates and inter-note links. In LightMem, the backend uses a multi-stage memory inspired by Atkinson–Shiffrin—sensory, short-term, and long-term—with efficient filtering; here, construction guidance prioritizes long-term promotion and retrieval uses stage-aware search (Lin et al., 19 Mar 2026).

The key interfaces are backend-agnostic. $M$6 translates MemMA’s atomic actions into backend-specific updates, and $M$7 calls the backend retrieval API while MemMA controls query formulation and evidence consolidation independently of the store (Lin et al., 19 Mar 2026). The paper further notes backend-specific preferences: Single-Agent benefits most from large $M$8 and more refinement rounds; LightMem is a stronger baseline and thus prefers moderate $M$9 and small $\pi_p$0 to avoid noise; A-Mem’s structure supports richer MERGE/INSERT decisions during consolidation (Lin et al., 19 Mar 2026).

Deployment requires no gradient-based training or reinforcement learning. MemMA is prompting-based: the Meta-Thinker, Memory Manager, and Query Reasoner are all LLMs, specifically GPT-4o-mini or Claude-Haiku-4.5, steered by carefully designed prompts (Lin et al., 19 Mar 2026). Prompt templates are structured: construction guidance extracts atomic facts and flags redundancy or conflict; answerability prompts enforce the blocking-gap test and output decision, reason, key-gaps, missing-speaker, time-need, and retrieval-guidance; repair prompts produce ADD_FACT with confidence and then SKIP/MERGE/INSERT with strict date and event rules (Lin et al., 19 Mar 2026). Dense self-supervision comes from probes, while LLM-as-a-Judge provides correctness labels for accepting or declining repairs.

6. Experimental evidence, ablations, and sensitivity

The primary evaluation is on LoCoMo, a benchmark for very long-term conversational memory. The reported evaluation subset is conv-26, consisting of 19 sessions, 419 turns, and 152 non-adversarial QA pairs across single-hop, multi-hop, temporal, and open-domain categories (Lin et al., 19 Mar 2026). Baselines are Full Text, Naive RAG, LangMem, A-Mem, and LightMem, with metrics F1 (token), BLEU-1, and ACC (LLM-as-a-Judge) (Lin et al., 19 Mar 2026).

With a GPT-4o-mini backbone, the reported overall LoCoMo results are: Full Text at F1 34.13, B1 24.63, ACC 61.18; Naive RAG at F1 27.14, B1 20.41, ACC 46.05; LangMem at F1 18.46, B1 14.05, ACC 30.26; A-Mem at F1 37.90, B1 28.85, ACC 52.63; LightMem at F1 44.58, B1 36.66, ACC 75.66; and MemMA_LM at F1 49.40, B1 38.28, ACC 81.58 (Lin et al., 19 Mar 2026). Category-level gains include Multi-Hop ACC improving from 65.62 to 78.12 and Single-Hop ACC from 78.57 to 82.86 when moving from LightMem to MemMA_LM (Lin et al., 19 Mar 2026).

Across backends, MemMA improves all three integrated stores in a plug-and-play manner. On GPT-4o-mini, Single-Agent improves from F1 22.64, B1 17.24, ACC 52.60 to MemMA_SA at F1 23.64, B1 12.94, ACC 84.87; A-Mem improves from F1 37.90, B1 28.85, ACC 52.63 to MemMA_AM at F1 46.23, B1 35.13, ACC 78.29; and LightMem improves from F1 44.58, B1 36.66, ACC 75.66 to MemMA_LM at F1 49.40, B1 38.28, ACC 81.58 (Lin et al., 19 Mar 2026). With a Claude-Haiku-4.5 backbone, LightMem reports F1 44.69, B1 37.77, ACC 73.03 and MemMA_LM reports F1 45.10, B1 36.53, ACC 76.97 (Lin et al., 19 Mar 2026).

Ablation results support the three principal components. On the Single-Agent backend, removing iterative retrieval reduces ACC from 84.87 to 70.39 under GPT-4o-mini and from 88.82 to 81.58 under Claude-Haiku-4.5; removing self-evolution reduces ACC from 84.87 to 73.68 under GPT-4o-mini; removing construction guidance reduces ACC from 88.82 to 83.55 under Claude-Haiku-4.5 (Lin et al., 19 Mar 2026). Budget sensitivity shows that MemMA_LM peaks at $\pi_p$1–$\pi_p$2 with ACC 81.58 and declines at $\pi_p$3, whereas MemMA_SA rises from 75.66 at $\pi_p$4 to 84.21 at $\pi_p$5. For refinement depth, MemMA_SA improves from 78.95 at $\pi_p$6 to 85.53 at $\pi_p$7, then falls to 81.58 at $\pi_p$8, indicating fast convergence and retrieval drift under excessive iterations (Lin et al., 19 Mar 2026). Probe quality also matters: on MemMA_LM with a Claude-Haiku-4.5 construction backbone, Haiku probes yield F1 44.98, B1 35.69, ACC 74.34; Sonnet probes yield F1 43.30, B1 32.74, ACC 74.34; Opus probes yield F1 45.10, B1 35.66, ACC 76.97, and the paper states that longer, cross-session probes yield stronger repairs than short keyword probes (Lin et al., 19 Mar 2026).

7. Distinctions, limitations, and reproducibility

MemMA should be distinguished from similarly named frameworks. “MMA: Multimodal Memory Agent” is a reliability- and abstention-aware multimodal memory agent that combines source credibility, temporal decay, and conflict-aware network consensus into a per-item reliability score and uses selective prediction; it is centered on multimodal reliability and abstention rather than planner–worker coordination of the memory cycle (Lu et al., 18 Feb 2026). “MEMA: Measurement Error in Meta-Analysis” is a Bayesian framework for continuous outcome data in statistical meta-analysis, addressing non-differential classical measurement error in covariates and individual participant data; it is unrelated to memory-augmented LLM agents despite the similar acronym (Campbell et al., 2020).

The limitations stated for MemMA are specific. It assumes that streams can be organized into sessions and that synthetic probes provide useful localized supervision, so less structured interactions may require adaptation. The reported evaluation is dialogue-centric, based on LoCoMo, and broader settings such as tools, web, or embodied agents require additional integration. The paper also raises ethical considerations: memory retention and repair must respect privacy, consent, and controllability, and automated repairs can risk persisting incorrect information without guardrails (Lin et al., 19 Mar 2026).

Reproducibility details are explicit. Code is available at the public repository listed by the paper. The key configuration uses GPT-4o-mini or Claude-Haiku-4.5 for the Meta-Thinker, Memory Manager, and Query Reasoner; GPT-4o-mini as Answer Agent and judge; text-embedding-3-small for embeddings; retrieval top-$\pi_p$9; refinement budget $\pi_s$0; and $\pi_s$1 probes per session, often generated by Claude-Opus-4.5 (Lin et al., 19 Mar 2026). Storage backends are Single-Agent, A-Mem, and LightMem, while Meta-Thinker prompts implement construction guidance plus answerability diagnosis with the blocking-gap test and anti-stall rule, and repair prompts implement ADD_FACT or NOOP followed by SKIP/MERGE/INSERT under strict date and event rules (Lin et al., 19 Mar 2026). Taken together, these details define MemMA as a prompting-based, backend-agnostic framework for coordinating construction, retrieval, verification, and repair across the memory cycle.