PaperlayerView exact revision on arXiv

Source-linked AI summary

MemoryLACE: Memory Lifecycle-Aware Consolidation and Evidence Retrieval

Meriem Yacoubi, Pia Schmidt, Nenad Petrovic, Ahmed Frikha, Martin Kirchhoff, Alois Knoll

arXiv:2609.03201v1cs.CLcs.LG

TL;DR

Long-term agents need to distinguish repeated evidence, updates, historical states, and unresolved contradictions, yet many textual systems leave these relationships implicit. MemLACE links local evidence lifecycles and retrieves related evidence together, achieving the strongest overall performance in same-backbone comparisons across BEAM and StructMemEval.

  • Problem

    Long-term agents need to preserve information across interactions, but textual memory systems often leave relationships among repeated, replaced, or contradictory evidence implicit.

    Source: §I · p. 1

  • Method

    MemLACE preserves atomic memories and provenance while using sparse lifecycle relations to reconstruct relation-aware evidence units for retrieval.

    Source: §I · p. 1

  • Results

    Across BEAM and StructMemEval, MemLACE achieves the strongest overall performance in same-backbone comparisons.

    Source: §VII · p. 8

  • Takeaways & Limitations

    The results indicate that modeling local evidence lifecycles can improve long-term memory reasoning without a comprehensive knowledge graph or global reflection.

    Source: §I · p. 1

Abstract

from arXiv · show

Long-term LLM agents must preserve information across interactions while distinguishing repeated evidence, historical states, updates, and unresolved contradictions. Existing textual memory systems retrieve semantically relevant memories efficiently but often leave these relationships implicit, whereas richer structured approaches model them through global graphs, hierarchical abstractions, or reflection at greater complexity. We introduce MemoryLACE (MemLACE), a lightweight memory framework that explicitly models the lifecycle of textual evidence through sparse merge, supersession, and contradiction relations while preserving atomic natural-language memories and their provenance. Rather than retrieving memories independently, MemLACE reconstructs relation-aware evidence units that expose current, historical, supporting, and conflicting evidence for downstream reasoning. Across BEAM and StructMemEval, using open-weight and proprietary LLM backbones, MemLACE achieves the highest overall performance in same-backbone comparisons while reducing end-to-end runtime on BEAM by 66.6% relative to Hindsight, the strongest reported reflective-memory baseline. Ablation studies identify lifecycle expansion and temporal awareness as the principal contributors to these gains. Together, the results demonstrate that explicitly modeling the local lifecycle of textual evidence is sufficient to substantially improve long-term memory reasoning without requiring comprehensive knowledge graphs or global reflection.

I. INTRODUCTION

Long-term agents need to preserve evolving evidence, but textual memory systems often leave relationships implicit while richer approaches model broader structure. MemLACE represents local evidence lifecycles and retrieves related evidence together.

I. INTRODUCTION

  • Long-term agents must preserve information across sessions, but textual systems often leave repeated, superseded, and contradictory evidence relationships implicit.The paper asks how agents can track changes and retrieve dependent memories without constructing a comprehensive global structure.Source: §I · p. 1
  • MemLACE preserves atomic memories and provenance while linking evidence through sparse merge, supersession, and contradiction relations for relation-aware retrieval.Its contributions also include lifecycle expansion and evaluation on BEAM and StructMemEval with open-weight and proprietary backbones.Source: §I · p. 1

II. RELATED WORK

  • Textual and event-level systems support long-horizon recall, but generally do not explicitly represent repetition, supersession, and unresolved contradiction as lifecycle relations.The cited examples include Mem0 [17], SimpleMem [5], SeCom [6], MemInsight, and EMem [8].Source: §II · pp. 1–2
  • Related systems model temporal continuity through broader semantic, causal, hierarchical, or reflective structures, whereas MemLACE limits relations to local evidence lifecycle links.The cited approaches include A-MEM, Zep [24], Theanine, TiMem [10], HiMem [11], Reflective Memory Management, Hindsight [15], and MemOS [18].Source: §II · p. 2
  • BEAM tests long-context conversational memory across ten capabilities, while StructMemEval tests whether accumulated evidence supports structured reasoning.The benchmarks address complementary settings targeted by the method.Source: §II · p. 2

III. MEMORYLACE ARCHITECTURE

MemLACE represents long-term memory as evolving atomic evidence with temporal metadata, provenance, activity state, and sparse lifecycle relations. Its retrieval pipeline uses this local structure without requiring a comprehensive knowledge graph.

III. MEMORYLACE ARCHITECTURE

  • MemLACE has a memory representation and two stages: construction and consolidation, followed by evidence retrieval and context construction.Retrieval begins with relevant active memories, expands their local relations, and organizes connected entries before answering.Source: §III · p. 2

A. Memory Representation

  • Each memory stores an atomic natural-language evidence restatement, temporal metadata, lightweight metadata, source provenance, activity, and lifecycle references.Lifecycle relations connect repeated, superseded, superseding, or conflicting memories; supersession links point in both directions.Source: A. Memory Representation · p. 2
  • The memory store is a sparse graph of textual memories linked by directed supersession edges and symmetric merge and contradiction links.It includes only relations needed to interpret stored evidence, keeping provenance distinct from lifecycle connections.Source: A. Memory Representation · p. 2

B. Stage I: Memory Construction and Consolidation

Stage I creates atomic memories from dialogue and consolidates them while preserving provenance. It then compares new evidence with a bounded, locally expanded candidate set to select lifecycle relations.

B. Stage I: Memory Construction and Consolidation

  • Overlapping dialogue windows are converted into standalone atomic memories, with duplicate entries introduced by overlap consolidated before lifecycle processing.Atomic entries capture one explicit fact, preference, plan, correction, decision, event, or state, preserving stated details and source-turn provenance.Source: B. Stage I: Memory Construction and… · pp. 2–3, 1) Memory Construction · p. 3

1) Memory Construction:

  • Atomic extraction collapses duplicate claims from the same turns and combines their provenance, while repetitions from distinct turns remain separate for possible merge linking.This granularity lets consolidation target an individual claim within a broader dialogue segment.Source: 1) Memory Construction · p. 3

2) Candidate and Relation Selection:

  • New memories are processed chronologically; similarity and entity matches are expanded to one-hop lifecycle neighbors before relation selection.The module chooses one strategy and, when applicable, one target; it distinguishes duplicate evidence, time-ordered state changes, incompatible same-scope claims, and independent memories.Source: 2) Candidate and Relation Selection · p. 3

3) Lifecycle State Update:

Lifecycle updates change which memories are active and how they connect, rather than deleting source evidence. Updates, unresolved contradictions, and merges receive distinct treatment.

3) Lifecycle State Update:

  • By default, lifecycle decisions change memory activity and connectivity without deleting source evidence.Source: 3) Lifecycle State Update · p. 3
  • When a newer memory updates an older one, the older entry becomes inactive and the newer entry becomes the active chain head.Source: 3) Lifecycle State Update · p. 3
  • Contradictory memories remain active and symmetrically linked; default linked merges retain original entries, while compact merges replace them with one synthesized memory.Superseded states are excluded from ordinary active retrieval, while historical, merge-linked, and conflicting evidence remains stored.Source: 3) Lifecycle State Update · p. 3

4) Local Batch Memory Evolution:

MEMLACE combines bounded, local lifecycle maintenance with hybrid retrieval over active memories. Retrieved anchors can recover inactive memories through their lifecycle relations when historical evidence is needed.

4) Local Batch Memory Evolution:

  • Local batch evolution reviews new memories and consolidation candidates jointly, adding update, contradiction, or merge links that become apparent only as a group.It changes relation links and activity states without rewriting memory text or adding factual content.Source: 4) Local Batch Memory Evolution · p. 3

C. Stage II: Evidence Retrieval and Context Construction

1) Hybrid Active-Memory Retrieval:

  • Hybrid retrieval fuses semantic, keyword, structured-template, and topic rankings with Reciprocal Rank Fusion over active memories.Inactive memories remain stored and can be recovered through retrieved anchors’ lifecycle relations when historical evidence is required.Source: 1) Hybrid Active-Memory Retrieval · pp. 3–4

2) Lifecycle Expansion and Evidence Units:

MEMLACE expands active retrieval anchors through bounded lifecycle relations, then groups related memories into evidence units. This exposes prior states and unresolved conflicts without traversing the full memory graph.

2) Lifecycle Expansion and Evidence Units:

  • Each active anchor expands along supersession lineages to prior states and one-hop contradiction links to directly conflicting evidence.The expansion remains local rather than traversing the full memory graph, limiting added query context.Source: 2) Lifecycle Expansion and Evidence… · p. 4
  • Retrieved memories form standalone or connected evidence units, with supersession chains ordered from earlier to newer states.This ordering keeps historical evidence and the active state visible together.Source: 2) Lifecycle Expansion and Evidence… · p. 4
  • Unresolved contradictory claims stay together, allowing relation-dependent memories to be reranked as coherent units rather than independent snippets.Source: 2) Lifecycle Expansion and Evidence… · p. 4

3) Chain-Aware Reranking and Context Packing:

MEMLACE reranks relation-connected evidence units and packs them within a context budget while preserving lifecycle order. The packed context distinguishes current, superseded, and conflicting memories for answer generation.

3) Chain-Aware Reranking and Context Packing:

  • Evidence units are reranked as groups, then selected under the context budget while preserving update-chain order and retaining complete units when possible.If a unit exceeds the remaining budget, only its higher-priority memories are included.Source: 3) Chain-Aware Reranking and Context… · p. 4

4) Evidence-Grounded Answer Generation:

IV. EXPERIMENTAL SETUP

  • The evaluation covers BEAM’s long-term conversational-memory benchmark and StructMemEval’s structured-reasoning tasks.Source: §IV · pp. 4–5

A. Benchmarks

On BEAM, MEMLACE outperforms all baselines on overall average at both model sizes, though its category-level gains are uneven. The evaluation also covers StructMemEval and uses backbone-matched comparisons.

A. Benchmarks

  • BEAM evaluates ten capabilities over long conversational histories; StructMemEval tests state tracking, tree-based, count-based, and recommendation reasoning.The study uses BEAM’s 100K-token subset and StructMemEval’s 51-scenario main set; BEAM comparisons include LIGHT, Hindsight, and SimpleMem, while StructMemEval compares with Mem-Agent.Source: A. Benchmarks · p. 5, §V · p. 5

B. Experimental Configuration

  • The study aligns MEMLACE with baseline model and evaluation settings, using Qwen3.5 backbones on BEAM and GPT-5.5 on StructMemEval.Generated answers are scored by GPT-4o mini as the fixed LLM judge.Source: B. Experimental Configuration · p. 5

A. BEAM Category-Level Comparison

  • MEMLACE substantially outperforms all baselines on BEAM overall average at both model sizes.Source: Table I · p. 6, A. BEAM Category-Level Comparison · p. 6
  • With Qwen3.5-4B, MEMLACE leads the 4B systems in abstention, contradiction resolution, event ordering, and instruction following.Source: A. BEAM Category-Level Comparison · p. 6
  • With Qwen3.5-9B, MEMLACE beats Hindsight in six of ten BEAM categories, with gains concentrated in tasks depending on several related memories.Knowledge update improves only marginally, where a single active-memory lookup suffices.Source: A. BEAM Category-Level Comparison · p. 6
  • Hindsight leads MEMLACE by 8.7 points in temporal reasoning, 5.2 in information extraction, and 4.9 in summarization.The paper links this trade-off to MEMLACE’s bounded, local evidence packing, which is less suited to exhaustive summarization and explicit temporal-interval calculation.Source: A. BEAM Category-Level Comparison · p. 6

1) Runtime Comparison:

On the BEAM 100K subset with Qwen3.5-9B, MemLACE uses less runtime than Hindsight, with its largest saving in memory construction.

Runtime Comparison

  • The comparison uses the BEAM 100K subset and Qwen3.5-9B under the same evaluation settings.The reported setup also specifies a vLLM serving environment, parallel processing, and one NVIDIA A40 virtual GPU profile with 24 GB allocated memory.Source: 1) Runtime Comparison · p. 6
  • Table II presents the runtime breakdown by phase for the BEAM 100K subset with Qwen3.5-9B.Source: Table II · p. 6
  • 66.6% lower total system runtime (2.99× speedup) and 3.02× faster memory construction are reported for MemLACE versus Hindsight.The largest saving occurs in memory construction, where MemLACE uses bounded consolidation and local batch repair instead of Hindsight's global reflection stage.Source: 1) Runtime Comparison · p. 6, Table II · p. 6

B. StructMemEval Results

On StructMemEval, MemLACE outperforms Mem-Agent by more than 17 percentage points, with perfect results on state-tracking and tree-based scenarios, while counting and recommendation remain difficult.

B. StructMemEval Results

  • MemLACE outperforms Mem-Agent by more than 17 percentage points, solving all 14 state-tracking and 10 tree-based scenarios versus Mem-Agent's 57.00% and 50.00%, with perfect accuracy lifting the macro-average to 52.08%.Source: B. StructMemEval Results · pp. 6–7
  • MemLACE reconstructs queried hierarchies from atomic relational memories and grouped evidence units rather than a comprehensive hierarchical knowledge graph.The passage attributes the gap over Mem-Agent entirely to the structured, evolving-state task types.Source: B. StructMemEval Results · p. 7
  • Neither method solves count-based scenarios; MemLACE reaches 8.33% scenario-level recommendation accuracy despite 32.2% correct individual questions.The recommendation scenario score uses the benchmark's 50% correctness threshold; the tasks depend on global numerical accumulation or ranking across distributed preference signals.Source: B. StructMemEval Results · p. 7

VI. ABLATION STUDY

The conclusion reports strongest overall performance in same-backbone comparisons across BEAM and StructMemEval, with lower runtime than Hindsight. Ablations identify lifecycle expansion and temporal awareness as the largest contributors, with distinct effects on retrieval and interpretation.

VI. ABLATION STUDY

  • Ablations evaluate one component change at a time on the complete BEAM 100K subset using Qwen3.5-9B, except for the combined-removal variant.Source: §VI · p. 7
  • Lifecycle expansion (−5.16) and temporal awareness (−4.97) account for the two largest reductions when removed, and affect distinct capabilities.Source: §VI · p. 7
  • Removing lifecycle expansion reduces knowledge update (51.9 →35.0), multi-hop reasoning (55.3 →42.3), and temporal reasoning (26.3 → 13.8).Source: §VI · p. 7
  • Removing temporal awareness reduces contradiction resolution (40.3 →26.6), preference following (84.4 →75.6), and temporal reasoning (26.3 →16.3).Source: §VI · p. 7
  • Compact merge improves contradiction resolution (40.3 →44.1) but reduces multi-hop reasoning (55.3 →43.0), yielding a net drop of 1.43 percentage points.The default linked-merge policy strikes a better trade-off by preserving separate provenance and phrasing.Source: §VI · p. 7
  • Removing batch evolution (−1.68) most affects instruction following (72.5 →60.6) and multi-hop reasoning (55.3 →48.2); removing both rerankers yields the smallest drop (−1.23).Batch evolution recovers relations spanning separately handled entries; rerankers reweight retrieved evidence.Source: §VI · p. 7
  • Combined removal drops the overall score to 43.3%, with subadditive reductions concentrated in evolving, relation-dependent categories.Event ordering (25.2) and summarization (35.9) remain close to the full configuration.Source: §VI · pp. 7–8

VII. CONCLUSION

  • Across BEAM and StructMemEval, MEMLACE achieves the strongest overall performance in same-backbone comparisons while reducing runtime relative to Hindsight.Source: §VII · p. 8

Related papers

Loading 2609.03201v1…