- The paper introduces ReFind, which uses an adaptive ReAct agent to search unmodified chat logs with BM25, session-aware ranking, context expansion, temporal filters, and deduplication.
- ReFind achieves 58.2% mean accuracy across six MemoryAgentBench tasks, outperforming structured systems such as HippoRAG 2 at 53.2% and leading on five benchmarks.
- The results suggest that faithful storage plus controllable, multi-round search can outperform costly offline memory construction while keeping evidence auditable and new conversations immediately searchable.
Motivation and central question
Most agent-memory systems impose structure on raw conversation histories before any question is asked—converting them into summaries, embeddings, trees, or knowledge graphs. The authors of ReFind observe that every such transformation is a bet placed in advance of the query: it decides what to abstract and what to discard, and omitted details may be unrecoverable later. They pose a question that the accumulation of increasingly elaborate memory architectures has largely left untested: how much of the reported benefit of structured memory comes from the structure itself, rather than from competent retrieval over the raw history? Their answer is to take the opposite extreme—build no semantic structure at all—and measure how far that setting can be pushed.
The design is grounded in empirical studies of personal-information refinding. Naturalistic work shows that users rarely rely on a single fully specified keyword query; they "orienteer" through small context-guided steps (Maharana et al., 2024), combining keyword trials with scrolling and temporal or contextual landmarks in instant messaging, and email field studies (345 users, over 85,000 refinding actions) find search and threading more effective than preparatory foldering. ReFind operationalizes this behavioral evidence as an agent-facing search interface over the unmodified record.
System design
ReFind separates retrieval from reasoning. In Stage 1, a ReAct-style controller issues parameterized BM25 queries over a turn-granularity inverted index of the raw archive, saves relevant fragments as notes, and reformulates based on observations; at most four iterations are allowed. In Stage 2, a separate answer model receives only the collected notes—grouped by session and sorted chronologically—so answer generation does not compete with retrieval for context space.
Crucially, "no structure" means no learned or LLM-generated memory representation. The system exploits organization already present in chat archives (turn boundaries, session IDs, timestamps) but performs no summarization, entity extraction, graph linking, or memory-card rewriting; indexing requires zero model calls and new turns are searchable immediately. Four chat-native controls complete the interface:
- Session-aware rank fusion: Reciprocal Rank Fusion combines turn-level BM25 ranks with session-level aggregate ranks (k=60), boosting sessions where multiple turns match.
- Context window expansion: each hit returns ±2 neighboring turns, truncated at session boundaries, restoring pronouns and ellipses.
- Temporal filtering: the agent supplies date ranges that prune candidates before scoring.
- Seen-session deduplication: sessions returned in prior rounds are excluded, the only stateful operation across rounds.
Conventional RAG is the degenerate case of this action space with one round and all controls inert; what the interface adds is not a better scoring function but state-dependent control over search.
Main results
On six MemoryAgentBench benchmarks under the incremental multi-turn setting—single-/multi-hop QA, LongMemEval, EventQA, and single-/multi-hop FactConsolidation, roughly 2,800 questions probing precise retrieval and factual-update tracking—ReFind attains the highest mean accuracy of any compared system with a GPT-4o-mini backbone matched to all reused baselines:
| Method |
SH-QA |
MH-QA |
LME |
EventQA |
FC-SH |
FC-MH |
Avg |
| GPT-4o-mini (long-context) |
64.0 |
43.0 |
30.7 |
59.0 |
45.0 |
5.0 |
41.1 |
| BM25-RAG |
66.0 |
56.0 |
45.3 |
74.6 |
48.0 |
3.0 |
48.8 |
| HippoRAG 2 |
76.0 |
66.0 |
50.7 |
67.6 |
54.0 |
5.0 |
53.2 |
| MIRIX |
62.0 |
61.0 |
37.3 |
29.8 |
14.0 |
2.0 |
34.4 |
| ReFind |
83.0 |
69.0 |
51.3 |
74.1 |
62.7 |
8.8 |
58.2 |
ReFind leads on five of six benchmarks. The gains over single-shot BM25-RAG concentrate precisely where one query is least likely to suffice: +17 points on single-hop QA and +13 on multi-hop QA, plus substantial lifts on both fact-consolidation tasks. On EventQA, ReFind and BM25-RAG are nearly tied, indicating that a strong lexical hit already solves many single-event questions. Notably, effective query formulation does not require a frontier model.
With a stronger GPT-5-mini controller on the LongMemEval-S/M subsets used by STITCH, the advantage widens: 93.2±3.3 (S) and 89.3±6.0 (M) over five runs, above STITCH itself (86.0/80.0), GraphRAG and HippoRAG 2 (both 66.7 on M), and GAM (70.0/60.0)—the closest design comparison, which also searches raw records but lacks the full chat-native interface. The margin over HippoRAG 2 grows from 0.6 points under GPT-4o-mini to 13.2/22.6 points under GPT-5-mini.
Ablations and attribution
Three matched control families rule out simpler explanations for the result:
- Generic agentic BM25 (same controller, budget, note-taking, and answer stage, but turn-level hits only): 78.7±4.6 / 82.2±3.8, i.e., −14.5/−7.1 versus the full method. Iteration alone does not explain the score; the four chat-native controls are jointly consequential.
- One search (agent-written first query, then all results passed directly to answering): −8.5 (S) and −20.4 (M). A well-formed initial query alone does not recover the result either, and the larger M gap is consistent with longer histories benefiting more from result-conditioned reformulation and skipping inspected sessions.
- Backend swap (dense text-embedding-3-large or four-way hybrid, everything else fixed): neither exceeds BM25's mean on either subset. Lexical matching is not the bottleneck—the agent supplies semantic adaptation through reformulation while the retrieval layer remains exact, transparent, and incrementally maintainable.
Component ablations attribute distinct roles: removing the context window causes the largest S drop (−9.2), since isolated lexical matches lack interpretable dialogue; removing session deduplication costs the most on M (−9.3), because longer histories otherwise repeatedly return the same sessions and exhaust the iteration budget; RRF reranking contributes several points on both subsets; temporal filtering matters least on these subsets but shows its value on fact-update tracking in the main suite.
Cost profile
Computation is deferred to question time: the full method averages 2.4–2.6 searches and ~5 LLM calls per query (~70–99K tokens per task). By contrast, the paper notes that GraphRAG extracted 8,564 entities and 20,691 relations in 281 minutes of GPT-4-turbo time on one ~1M-token corpus before answering anything. ReFind thus ties model computation to expressed information needs rather than to every archive update, and answers remain auditable against verbatim turns rather than latent memory states.
Limitations and open questions
The paper is explicit about scope. The claimed advantage covers precise retrieval and factual-update tracking; the authors position ReFind as complementary to semantic-abstraction and low-latency memory mechanisms rather than a replacement for them. Several caveats qualify the numbers: baseline scores on the main suite are reused from prior papers rather than reproduced locally, and systems were not all rerun with a matched controller or tool-call budget; the LongMemEval-S subset contains only 50 questions and M only 15, so a single changed judgment moves M accuracy by 6.7 points, making run-level dispersion essential to interpretation; deltas between conditions with different repetition counts are differences of observed means, not paired estimates; and no independent human-agreement study was conducted for the LLM judges. Temporal filtering showed little effect on the LongMemEval subsets due to their low proportion of time-sensitive questions, so its contribution rests largely on the consolidation tasks. Open questions include how the interface performs on workloads demanding abstraction or strict latency guarantees, and which derived structures justify their offline cost once faithful storage plus controllable search is the default.
Conclusion
ReFind demonstrates that leaving a conversational archive unmodified and exposing its native structure—sessions, timestamps, neighborhoods, visited-state—to an adaptive agent outperforms transforming it into a memory representation first, achieving 58.2 mean accuracy against 53.2 for the strongest structured baseline over 2,800 questions, and 93.2±3.3/89.3±6.0 on LongMemEval-S/M with a stronger backbone. Controlled comparisons attribute the gain to the interaction between multi-round agent control and a conversationally structured lexical interface, not to model scale, semantic embeddings, or a lucky first query. The practical implication is a modular design principle: begin with faithful storage and a controllable search interface, adding derived structures only where a specific workload demands them.