代理記憶與檢索
ReFind Skips Semantic Memory Indexes, Beats HippoRAG 2 by Five Points on Average Using Raw Chat-Log Search
ReFind preserves conversations without rewriting them, allowing an agent to repeatedly gather evidence using keyword, time, and session filters. Across roughly 2,800 test questions, it achieved an average accuracy of 58.2%, surpassing the strongest structured-memory baseline at 53.2%, though the comparison did not standardize all controllers and tool-call budgets.
Most long-term memory systems first transform conversations into summaries, vectors, trees, or knowledge graphs, then wait for future queries. A new study proposes the opposite approach with ReFind: raw logs receive only a lexical index organized by conversation turn, with no embeddings or semantic graphs generated. A ReAct agent can reformulate keywords over multiple rounds and use four controls—session-aware rank fusion, adjacent-context expansion, time-range narrowing, and skipping previously inspected sessions—before passing the collected evidence to a separate reasoning stage to produce an answer.
The researchers evaluated about 2,800 questions across six MemoryAgentBench task groups, including single-hop, multi-hop, event-ordering, and fact-update tasks. Using the same GPT-4o-mini backbone, ReFind achieved an unweighted mean accuracy of 58.2%, compared with 53.2% for HippoRAG 2 and 48.8% for one-shot BM25-RAG. On multi-hop QA, ReFind scored 69.0%, versus 56.0% for BM25. On EventQA, however, ReFind scored 74.1%, slightly below BM25’s 74.6%, suggesting that agentic search may be unnecessary for a single, clearly defined event.
The engineering implication is not that knowledge graphs are obsolete, but that “compress first, retrieve later” may discard details too early. ReFind incurs no LLM-based indexing cost, and new messages become searchable immediately. The tradeoff is that costs shift to multiple model calls for each query, and the paper does not provide a complete latency and cost comparison. More importantly, some baseline figures were taken from prior research, and the authors acknowledge that the systems were not rerun under fully identical controllers and tool-call budgets. Some LongMemEval results also rely on a GPT-4.1-mini judge, without a human agreement study. The next things to watch are a public implementation, end-to-end costs, and whether lexical search can retain its advantage on semantically ambiguous questions rather than precise factual ones.