Agent memory systems finish a run, compress it into a lesson, and throw the run away — deciding what mattered before anyone knows what the next task will be. Keep the transcripts instead and do the compressing when a task arrives, and a small executor goes from 61.2% to 77.4% on household tasks and 16.5% to 32.8% on shopping, while the memory it carries into context shrinks from 10.7–13.4K tokens to 1.9K. The note is short because it is aimed.
Every agent memory system you can install works the same way. A run finishes, a model summarises it into something reusable — a skill, a heuristic, a lesson — and the transcript is dropped. The summary is cheap to retrieve and cheap to carry, and that is the point.
It is also a decision made in the dark. Whatever is thrown away is gone irreversibly, and whatever is kept has to serve every future task at once. A single run through a kitchen can teach one task that the sink has to be empty before you can wash anything, and teach a different task that mugs live in the cabinet rather than on the counter. Which of those is the lesson depends on a question nobody has asked yet.
Pick an incoming task and switch the note.
Nothing here is a new retrieval algorithm. The store is plain keyword search over task descriptions, and the trajectories it returns are whole and unedited. The only structural change is where the summarising happens: after the run, or after the question.
The payoff shows up twice. The note is written for one task, so it is shorter, and it is written after the task is known, so the reward for a bad note arrives immediately instead of several tasks later — which is what makes the curator trainable at all.
Three benchmarks: a text household simulator, a shopping environment, and a customer-service setting with tool calls and a simulated user. Three executors, from an 8B open model to two frontier ones. Switch between them.
The untrained version is the interesting one. With an 8B executor it is no better than write-time curation on shopping — 11.7% against 13.6% — but hand the same setup a frontier executor and read-time curation wins without any training at all: 79.3% against 69.8% on household tasks, 39.3% against 33.6% on shopping. The stronger the model doing the reading, the more there is to gain from not having thrown the transcript away.
Each of these removes one piece and re-measures. They are the answer to the obvious objection — that this is just retrieval with extra steps — because the two biggest drops come from removing the retrieved transcripts and from hiding the task from the curator, which are exactly the two halves of the idea.
The success numbers are the headline, but the context bill is the part you will feel first. A memory system that prepends thirteen thousand tokens to every task is a real line on an invoice and a real slice of a context window. Set your volume and your input price.