h t t p s : / / a r x i v . o r g / a b s / 2 6 0 9 . 0 3 2 0 1

Ingest: MemoryLACE — Memory Lifecycle-Aware Consolidation and Evidence Retrieval

Type: kb/sources/types/ingest-report.md

Classification

Scientific paper: an arXiv v1 preprint by Meriem Yacoubi and colleagues, affiliated with the Technical University of Munich, inovex, and Cerebras. The authors propose MemoryLACE, report benchmark comparisons and component ablations, and interpret their own system's results. This is experimental design evidence, not an independent replication or a production-use report.

Summary

MemoryLACE retains atomic natural-language memories with source-turn provenance and connects repeated, superseded, and contradictory entries. New evidence changes activity states and relations; retrieval begins with active entries, expands historical and conflicting evidence, then ranks and packs connected groups for answering. On BEAM's 100K subset, the reported overall scores are 45.4% with Qwen3.5-4B and 51.9% with Qwen3.5-9B, versus Hindsight's 40.1% and 50.3%; the 9B run takes 66.6% less total time in the tested serving environment. Within MemoryLACE's atomic-memory pipeline, removing lifecycle expansion or temporal awareness produces the largest ablation losses. Replacing linked entries with a synthesized memory improves contradiction resolution but worsens multi-hop reasoning. StructMemEval results likewise favor state and tree tasks while exposing poor counting and recommendation performance. The paper supports selective retrieval of retained evidence and operational lifecycle state within this setup, without establishing that sparse lifecycle relations generally outperform richer representations.

Quotes

No source quotes have been retained yet.

Connections Found

The paper supplies a bounded empirical case for Retire, Redact, Supersede, And Relax Memory: superseded entries leave ordinary active retrieval, but remain recoverable through their successors. Temporal metadata helps distinguish an ordered update from a same-scope conflict. The tested mechanism addresses supersession and temporal validity, not the note's redaction, retirement, or policy-relaxation requirements.

It also supports the capture/loading distinction in Preserve Evidence Without Making History The Next Context. The default merge policy retains separate extracted memories and provenance while query-time selection controls what enters context. In the fixed atomic-memory pipeline, compact merging trades better contradiction resolution for worse multi-hop performance. This is evidence for preserving distinctions that later questions may need; it does not test comprehensive raw-trace retention or establish that all compression loses useful evidence.

Learning Claims (our opinion)

MemoryLACE adapts external memory to incoming dialogue. Its consolidation model chooses merge, update, contradict, or no action against bounded candidates. An update makes the older entry inactive; an unresolved contradiction leaves both entries active; the default merge links repeated evidence without synthesizing it. A later local batch pass repairs relations and activity states without rewriting the extracted text or adding facts. Answering consumes the resulting state, so these changes have an explicit route to later behavior.

This maps to retained evidence maintenance and context engineering. It does not establish theory refinement: the demonstrated operation classifies relationships between explicit dialogue claims, rather than deriving a theory's consequences, diagnosing a failed prediction, and revising the responsible theory. The experiments use a fixed backbone and supply no demonstrated feedback loop that learns a better consolidation policy from answer errors.

The boundary described in Learning inside a fixed decomposition inherits its mistakes matters here. Atomic extraction, the relation vocabulary, bounded candidate search, and the retrieval-to-answer interface are supplied by the designers. Memory evolution changes entries and links within that arrangement. The ablations test particular choices inside it, including linked versus compact merge, but do not search alternative memory granularities or operation sets. Poor counting performance identifies an observed weakness; it does not prove counting is inexpressible through the model's available reasoning. Dedicated aggregation remains a proposed extension rather than a tested remedy.

Extractable Value

  1. Preserving individual evidence has a task-dependent payoff. In the Qwen3.5-9B BEAM ablation, compact merge raises contradiction resolution from 40.3% to 44.1% but lowers multi-hop reasoning from 55.3% to 43.0%, with a reported overall loss of 1.43 percentage points. This sharpens the evidence-preservation requirement: test what synthesis removes against later uses, rather than treating consolidation as uniformly beneficial or harmful. The contrast varies the merge policy within one atomic-memory pipeline; it does not isolate provenance loss from changed wording, memory count, or retrieval behavior. [quick-win]

  2. Retrieving history and interpreting time solve different problems. Removing lifecycle expansion lowers the overall score by 5.16 points; removing temporal awareness lowers it by 4.97. Expansion removal strongly affects updates and multi-hop questions but leaves contradiction resolution unchanged at reported precision, whereas temporal-awareness removal lowers contradiction resolution from 40.3% to 26.6%. Within this pipeline, relation traversal and evidence-grounded temporal interpretation make distinct contributions. A KB evaluation should therefore check both whether the needed versions arrive and whether their temporal scope is interpreted correctly. [experiment]

  3. Connected evidence groups are a concrete retrieval design to test. Start from active entries, recover bounded supersession lineages and conflicts, then rank related entries jointly. This gives the existing retention/loading distinction an operational comparison target: independent snippets versus connected groups under a controlled context budget. The paper does not separately ablate every grouping and packing choice, so it motivates that comparison without settling it. [experiment]

Limitations (our opinion)

The evaluation covers 20 conversations and 400 questions from BEAM's 100K subset, plus 51 StructMemEval scenarios. It uses GPT-4o mini as the fixed answer judge and reports no repeated-run uncertainty or statistical significance. Small overall differences, especially the 1.6-point 9B advantage over Hindsight, should not become general superiority claims. StructMemEval's 100% state and tree figures count solved scenarios, with a scenario solved when at least half its questions are correct; they do not establish perfect question-level accuracy. Counting remains at zero solved scenarios, and recommendation reaches only 8.33%.

The component ablations hold most of the supplied decomposition fixed. Comparisons against other systems vary several mechanisms together, and sharing SimpleMem's construction scheme does not isolate every lifecycle design choice. The compact-merge experiment changes a bundle of text retention, provenance, and retrieval conditions; the authors' explanation in terms of lost evidence is plausible but not a separately identified cause. Likewise, the unchanged contradiction score without expansion does not itself demonstrate that the relevant conflicts were independently retrieved.

The runtime result belongs to the reported Qwen3.5-9B, vLLM, parallel-processing, and 24-GB A40 virtual-GPU setting. No implementation was inspected or executed for this ingest, and the paper's performance and configuration claims remain unreproduced. The source specifies bounded expansion and a 64-memory final context, which can omit historical evidence or split a group when the budget is exhausted; scaling, extraction mistakes, and relation errors need further evaluation. Retained atomic restatements with provenance are not equivalent to complete source-history retention. The paper does not test Commonplace's theory revision, governance, or lifecycle requirements beyond conversational and structured-memory tasks.

Update Preserve Evidence Without Making History The Next Context with the linked-versus-compact merge result as a bounded empirical example of task-dependent evidence loss during consolidation, retaining the fixed-pipeline and mixed-outcome qualifications.