Your saved summary is an index, not evidence

Overview

Three weeks ago you had an assistant pull the key points out of a long thread and keep them. Today you ask the question that actually matters — can we commit to that date, is that constraint still live — and it answers from the note. Not from the thread. From the note.

That feels like the disciplined version: there is a written record, and a record seems more accountable than trusting a chat to remember. A February 2026 paper measured the same system answering later questions both ways, and the tidy version is the worse one. Answering from its own saved summaries scored 0.755; answering from the raw source it had also kept scored 0.873. On a second benchmark, 0.678 against 0.808. Twelve to thirteen points of accuracy, and the only thing that changed was which copy the answer came from.

The habit this builds: when a saved summary or project memory is about to carry a decision, make it name the source file and quote the line — then open that file yourself before you act.

The content

The researchers named the mechanism, and the name is worth keeping: the write-before-query barrier. Whatever the note kept, it kept for reasons that had nothing to do with today’s decision, because today’s decision had not been made yet. So it cannot contain what it was never asked to keep, and — this is the part that bites — it cannot tell you that it doesn’t.

Their own illustration is the clearest version: compressing a “severe peanut allergy” into generic “dietary preferences” creates an unverifiable omission. Nothing in that note is false. It simply cannot answer is this snack safe, and will not say so.

Now overturn the obvious read. Saving the summary looks like the move that makes a decision more defensible — you can point at what informed it. But defensibility and accuracy pull in opposite directions here. The record is genuine. It is just not the evidence, and the moment it stands in for the evidence, the decision quietly gets worse while looking better.

This is not a quirk of memory features — the same shape turns up wherever a system abstracts before it answers. At EMNLP 2025, a Stanford team compared plain retrieve-then-read — pull the relevant passages, keep the document’s original order, read them — against more elaborate pipelines that summarise first. The plain method “consistently matches or outperforms more intricate methods”, which they trace to keeping source fidelity and structure intact rather than to anything clever. The abstraction step did not pay for itself.

A different lab pointed the same way in June 2026, building an agent memory that retains verbatim source excerpts rather than write-time summaries; at a matched budget it beat the strongest comparable alternative. That is the independent corroboration the February paper cannot supply alone.

Two honest boundaries, and both change how you apply this.

The first: the answer is not paste everything. This site has covered why long, indiscriminate context degrades an answer, and that stands. The February 2026 paper concedes it directly — raw source by default is expensive, and its fix routes to the raw store only when the summary is insufficient, recovering most of the accuracy on one benchmark at 54.1% fewer input tokens. So the move is reattach the relevant source, never dump the corpus.

The second is the closest thing to a human test, and it cuts the other way. In a 2025 study, physicians reviewing heart-failure admissions with an AI summary did no better and no worse than working from the chart alone. No harm. But look at the configuration: the summary was additive, the chart stayed open beside it, and clinicians went back to it when the summary looked thin. That is keep the source in the room, and it was safe. It is not evidence that deciding from the note alone is safe, because nobody there had to.

Which gives you the distinction to carry, and the line against the write-step module this site already ran: memory and saved summaries are the index. The connected files are the evidence. The turn where you decide needs the evidence in it.

Try it

Add this to the prompt you use before any decision an assistant helped you form. It needs the source connected or uploaded, so check what your workspace has enabled — if connectors are off, upload the file; in plain chat, the same tagging works against earlier turns.

Answer only from the files attached or connected to this chat.
If any part of your answer comes from project memory, a saved
summary, or an earlier turn, mark it [MEMORY], name the source
file, and quote the sentence I should check it against.

Then the human half, which is the actual module: if it cannot give you a filename and a sentence you could look up, treat that recall as an index entry, not a finding — and open the file. Reattach the section you need, not the whole corpus.

Where it breaks. The tag flags what came from the note; it cannot flag what the note left out, and no prompt recovers a sentence nobody wrote down. And if you will not open the file, the tagging is worse than useless: it makes a thinner answer look better sourced.

Additional reading

• From Lossy to Verified: A Provenance-Aware Tiered Memory for Agents — Zhu et al., February 2026. Source of the write-before-query barrier and the summary-versus-raw comparison (LoCoMo 0.755 vs 0.873; LongMemEval 0.678 vs 0.808). The 54.1% routing result is LoCoMo-only — on LongMemEval the router recovers less of the gap (0.752 against 0.808). A workshop paper (ICLR 2026 MemAgents), not a main-conference result, and one team measuring its own system.
• Stronger Baselines for Retrieval-Augmented Generation with Long-Context Language Models — Laitenberger, Manning & Liu, EMNLP 2025. The peer-reviewed leg: retrieve-then-read preserving original passage order matches or beats pipelines that abstract first.
• EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents — Li, Banerjee & Che, June 2026. The independent leg: a different group storing evidence capsules — verbatim source excerpts with retrieval keys — rather than summaries, and beating the strongest comparable budgeted baseline. Still a preprint.
• Verifiable Summarization of Electronic Health Records — Verma et al., medRxiv preprint, June 2025. The counterweight, and small (8 physicians, 8 cases each): a summary alongside the chart was no worse than the chart alone. Clinicians kept full access and used it.

Editor’s note

I built my own personal memory system around this principle: verbatim recall underneath contextualised recall. It works very well, because it’s aimed at how models actually draw from context. Most enterprise harnesses won’t let you bolt on a memory layer of your own, so what’s left is the move in the module: keep the source in the room as an addition to the summary. It’s both easy to do, and easy to forget.