A Memory Index Has to Earn Its Rewrite

We often treat a pile of summaries, embeddings, and graphs as the natural price of agent memory. A new result says that price is not automatically buying the thing we want: grounded recall from records we can still inspect.

The signal

When Your Agent Opens the Chat App: Agent-Controlled Search over Raw Chat Logs Rivals Structured Memory asks a pleasantly rude question: how much of memory quality comes from the structure itself, and how much comes from competent retrieval? Its ReFind system leaves the chat archive unchanged, indexes turns lexically, then lets an agent search iteratively with session-aware ranking, local context expansion, temporal narrowing, and a way to skip sessions it has already inspected.

On roughly 2,800 MemoryAgentBench questions, ReFind reached mean accuracy of 58.2 using GPT-4o-mini. The strongest listed graph and tree baseline, HippoRAG 2, reached 53.2 under the same backbone. On LongMemEval-S/M, the paper reports 93.2 plus or minus 3.3 and 89.3 plus or minus 6.0 with GPT-5-mini.

The evidence

The interesting part is not simply that lexical search won one comparison. The authors also compare ReFind with single-shot BM25, a matched generic agentic BM25 control, component removals, and agentic dense and hybrid variants. That makes the claim narrower and more useful: the combination of agent control, chat-specific search controls, and unchanged records can recover much of the value often credited to elaborate memory structure.

For WTK, that is a direct challenge to the reflex that every memory problem needs another derived representation. A source-preserving record gives us a clearer provenance trail and a better way to inspect why the agent did or did not find the evidence it needed. Our take: a memory transformation should earn its retention by improving a declared task while preserving the evidence path needed to check the answer.

The boundary

This is a preprint about conversational-memory benchmarks, not a result about WTK packages, WTK memory assets, non-chat sources, or deployed outcomes. It does not prove that lexical retrieval is sufficient for every grounding obligation, that derived memory is harmful, or that raw records are safe to expose without access controls.

It also does not say that one model or one retrieval scheme is generally better. The reported values belong to the named benchmarks, controls, and model backbones. A transformed representation may still be worth its cost when a defined task and access policy show a benefit that a source-preserving baseline cannot match.

The builder impact

Builders should make the baseline difficult to dismiss. Before adding a summary tree or semantic graph, compare it with a controllable search path over the permitted source records. Keep the source links, access decision, query trace, retrieved evidence, answer, and failure cases. Then ask separately whether the added structure improves recall, changes the errors, or merely makes the system harder to inspect.

The WTK test

WTK can turn this into a bounded qualification fixture without importing the paper's archive or declaring a new memory subsystem. Freeze a small, permission-bounded record set and a grounded question set. Compare a source-preserving retrieval path with one proposed derived-memory projection under the same package, model, tool authority, and target. Retain every outcome and test both supported answers and justified unresolved results.

That would tell us whether a particular transformation earns its complexity in one stated execution form. Until then, this paper is a useful warning, not a WTK Finding and not a reason to rewrite memory by default.