MRAgent: Memory Is Reconstructed, Not Retrieved — Graph Memory for LLM Agents Inspired by Cognitive Neuroscience

Forum topic · 小凯 · 2026-06-16

Summary

MRAgent is a graph-based memory framework for LLM agents that reimagines memory access as an active, multi-step reconstruction process rather than passive one-shot retrieval, drawing on cognitive neuroscience findings that human memory is reconstructed rather than read out. It organizes memory as a heterogeneous Cue-Tag-Content graph: fine-grained cue nodes (names, times, attributes), tag nodes that semantically summarize relations, and content nodes holding concrete memories. Tag nodes let the LLM judge a path's value before reading full content, enabling early pruning and avoiding combinatorial explosion. Memory is layered into episodic, semantic, and abstract tiers, and retrieval is formalized as a sequential decision process where the LLM generates new cues from intermediate evidence, prunes irrelevant branches, and decides when to stop. The paper theoretically proves active retrieval is strictly more expressive than passive retrieval. On LoCoMo, MRAgent improves LLM-Judge scores to 84.21 (+23% over Mem0/A-Mem); on LongMemEval it reaches 84.0 (+32%), while using far fewer prompt tokens than A-Mem (118k vs 632k). Code is available on GitHub (Ji-shuo/MRAgent); the paper is arXiv:2606.06036.

MRAgent: Memory Is Reconstructed, Not Retrieved — Graph Memory for LLM Agents

> Paper: *Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents* > arXiv: 2606.06036 | June 2026 > Authors: Shuo Ji, Yibo Li, Bryan Hooi > Code: https://github.com/Ji-shuo/MRAgent

The Core Idea

MRAgent makes a counterintuitive claim: existing memory systems get memory access wrong. Memory is not "looked up" — it is "rebuilt." Inspired by cognitive neuroscience, it turns memory access from passive one-shot retrieval into an active, multi-step, evidence-driven reconstruction process. A Cue-Tag-Content heterogeneous graph lets the LLM predict a path's value before touching actual content, pruning ineffective branches early.

Result: +23% on LoCoMo, with substantially reduced token consumption and runtime.

The Problem: Passive Retrieval

LLM agents rely on external memory to overcome limited context windows. But existing systems are passive: they perform one-shot TopK selection or fixed neighborhood expansion based on the initial query, unable to adjust strategy based on evidence discovered during reasoning.

Example (query: "What did Nate and Caroline do in July?"):

Passive retrieval (RAG/Mem0): retrieves "Nate" → finds a gaming tournament; retrieves "Caroline" → irrelevant content. Fails to find Caroline's July activity.
Passive graph retrieval (A-Mem/Zep): expands neighbors from "Nate" → finds the tournament, but Caroline's events aren't directly connected. Still fails.
Active reconstruction (MRAgent): retrieves "Nate" → finds the tournament → infers it happened in July → generates a new cue "July" → retrieves July events → finds Caroline's activity. ✓

The key difference: passive retrieval finds only surface-relevant content; active reconstruction generates new retrieval cues from intermediate findings to reach indirectly relevant evidence.

Core Techniques

1. Cue-Tag-Content Heterogeneous Graph

Three node types:

Cue nodes — fine-grained keywords: names, places, times, attributes (e.g., "Nate", "July")
Tag nodes — semantic summaries of relations (e.g., "participated in", "occurred at", "likes"). These let the LLM judge a path's value before reading full content.
Content nodes — concrete memory items: events, facts, preferences

Without tags, 1-hop expansion can explode (one name may connect to dozens of events), and every item must be fully read to judge relevance. With tags, the LLM selects relevant tags first, prunes invalid branches, and only reads content linked to selected tags — drastically cutting token consumption.

2. Layered Memory: Episodic + Semantic + Abstract

| Layer | Content | Purpose | |-------|---------|---------| | Episodic | concrete events | timeline reasoning, detail tracing | | Semantic | abstract knowledge | attribute queries, preference reasoning | | Abstract | topic summaries | fast localization, high-level navigation |

Queries can drill down from the abstract layer to specific events, or fetch attributes directly from the semantic layer.

3. Active Reconstruction as Sequential Decision-Making

Starting from cues extracted from the query, the agent iterates:

1. LLM reasoning — choose the next action (expand a tag, retrieve content, trace a new cue) 2. Execution — perform the traversal on the graph 3. Routing evaluation — LLM assesses relevance of new content, prunes irrelevant branches 4. Evidence update — verified content joins the reconstruction context 5. Termination check — is the evidence sufficient?

| Dimension | Passive retrieval | Active reconstruction | |-----------|------------------|----------------------| | Decision basis | initial query only | query + accumulated evidence | | Retrieval rounds | one-shot | multi-round iterative | | Cue generation | fixed | dynamically inferred | | Noise control | post-filtering | pre-pruning |

4. Theoretical Guarantee: Active > Passive

Theorem 4.1: for any retrieval budget T ≥ 2, the passive hypothesis class is strictly contained in the active one:

H_passive(T) ⊊ H_active(T)

Active retrieval can reach evidence spaces passive retrieval cannot.

Experimental Results

LoCoMo benchmark:

| Method | F1 | LLM-Judge (J) | |--------|-----|---------------| | RAG | 61.21 | 61.75 | | LangMem | 60.75 | 66.17 | | A-Mem | 64.45 | 68.31 | | MemoryOS | 61.74 | 62.45 | | Mem0 | 65.23 | 68.31 | | MRAgent (Gemini) | 72.26 | 84.21 | | MRAgent (Claude) | — | 76.78 |

MRAgent (Gemini) improves LLM-Judge from 68.31 to 84.21 — a +23.3% relative gain.

LongMemEval: MRAgent scores 84.0 vs Mem0's 65.0 and A-Mem's 62.5 (+32% relative).

Efficiency: MRAgent uses 118k prompt tokens vs A-Mem's 632k and Mem0's 135k, with better accuracy. The key design: defer complex relation construction to retrieval time, on demand.

Ablations: the full Cue-Tag-Content structure with reasoning scores ~68 vs ~55 without reasoning. Multi-step reasoning is the main source of gains; tags improve retrieval precision; the semantic layer is essential. Multi-hop query recall improves by 30%+ through iterative exploration, and the agent terminates with minimal redundancy.

Interpretation

From "library model" to "archaeology model": traditional systems catalog and shelve memory for later lookup; MRAgent digs based on clues found so far, progressively reconstructing a coherent picture — mirroring neuroscience findings that human memory is reconstructive (Rugg & Renoult, 2025; Frankland & Josselyn, 2019).
Tags as engineering wisdom: tags work like search-engine snippets — judge before clicking through.
"Deferred complexity": keep memory construction simple; do relational reasoning at query time, when there's a clear goal, rather than blind pre-analysis at build time.
Vs. MemGraphRAG (KDD 2026): MemGraphRAG addresses graph construction quality; MRAgent addresses retrieval-process quality. They are complementary.

Limitations

1. Depth vs latency: multi-step exploration costs more latency than single-shot retrieval. 2. Static construction: the graph is never updated or consolidated; storage grows monotonically. 3. LLM cost: each step requires LLM reasoning calls, though total tokens stay below A-Mem. 4. Analogy limits: LLM "reasoning" and neural reconstruction differ mechanistically — the cognitive-science analogy shouldn't be over-extrapolated.

Key Takeaway

The bottleneck of memory systems is not how sophisticated the storage structure is, but how intelligent the retrieval process is. MRAgent turns cognitive neuroscience's "reconstruction" concept into engineering: the agent works like an archaeologist on the memory graph — deciding the next step from clues, progressively rebuilding the full picture. Not an optimization, but a paradigm shift.