From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory
MemCompile compiles query-conditioned evidence for long-term agent memory, reaching 80.52% judge accuracy on LoCoMo.
Researchers propose MemCompile, a query-conditioned evidence compilation framework for long-term conversational agent memory. An Address Plane locates support through source-linked descriptors and capacity-aware Cells, then an Evidence Plane allocates a bounded budget and renders selected claims into answer-ready context. MemCompile reports 80.52% judge accuracy on LoCoMo and strong results on LongMemEval-S, extending the quality-cost Pareto frontier.
- Separates memory addressing from how evidence is presented to the reader.
- Address Plane uses source-linked descriptors and capacity-aware Cells.
- Evidence Plane budgets complementary packets into answer-ready context.
- Reports 80.52% judge accuracy on LoCoMo and gains on LongMemEval-S.
From the original article
1 min read at arxiv.orgLong-term agent memory requires recovering answer-relevant evidence dispersed across turns and sessions. Existing systems increasingly use structured memories and adaptive retrieval, yet often pass retrieved or expanded units directly to the reader, coupling memory addressing with evidence presentation and leaving the final context redundant or incomplete.
By Shangze Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.