Reproduce “EdgeMem: LLM-Free Agent Memory Construction and Retrieval via Evidence-Preserving Multi-Anchor Hypergraph” on OpenResearch

About this paper

Zeyang Cui, Jiannong Cao, Zhiyuan Wen, Bo Yuan, Junlan Feng, Shengyuan Chen

EdgeMem builds long-term memory for LLM agents without using generative models during memory writing or retrieval. By preserving original conversation turns and linking them through content, time, and session anchors in a hypergraph, it achieves stronger retrieval and answer-quality metrics than reproduced baselines while reducing construction to zero LLM calls and reaching 10.73 ms P95 retrieval latency.

Key insights

• Episode-level retrieval provides the largest improvement over flat BM25 by restoring neighboring turns and broader conversational context; cross-session co-occurrence links add further gains.
• Keeping source evidence intact supports conservative responses to unsupported-premise questions, though the evidence-only policy also causes some abstention on answerable queries.
• Deterministic, append-only memory organization offers predictable cost and provenance, but performance remains dependent on entity extraction, lexical overlap, and reliable date metadata.

Summary by alphaXiv.