Persistent memory is arriving in production agent platforms: when a task closes, a second model reads the transcript and writes down what should carry to the next session. Working from the transcript alone, that writer inherits every one of the run's mistakes — it stores the answer instead of the procedure, promotes one observation into a general rule, and keeps schema names that were deleted last month. Give it a least-privilege, read-only copy of the tools the agent had, and make it check each note before committing: pass rate on a database-exploration set goes from 39% to 73%, the working agent's queries fall from 8.8 to 4.7 per question, and its cost falls from $3.38 to $1.68. The agent, the retriever, the memory format and the write permissions are all unchanged.
The shape is now standard. A task finishes; an asynchronous curator reads the trajectory, the terminal feedback and any nearby existing records, and writes durable notes for next time. Each note is small and structured: a category — pattern, rule, trap, schema, policy or interaction — a confidence, a short retrieval scope, and one actionable claim, plus provenance and usage metadata. The instruction to the curator is to store reusable facts, procedures and relations rather than task answers.
The problem is the evidence available to it. The trajectory shows what the agent did and how it was scored; it does not show whether a shortcut existed, whether a relation holds outside the one slice that was queried, or whether the environment has moved since. So the curator writes confidently about things it cannot see. Step through three records the paper found, and switch on the probe.
All three notes are well formed. All three would be retrieved next session, and two of them would actively mislead the agent that retrieved them — the trap tells it what not to answer without supplying the query that works, and the schema record points at a column that no longer exists. This is the failure the paper is aimed at: a curator restricted to completed trajectories preserves errors, overgeneralises partial evidence, and retains stale knowledge.
What that costs is visible in the headline comparison. Memory written from transcripts does beat having no memory at all — but it is beaten on reward by simply pasting every prior transcript into the context window, and its spread across runs is enormous.
Environment-probing curation adds one phase between writing a note and keeping it. The curator still proposes atomic candidate memories from the trajectory and the feedback. Before committing, it is allowed to make targeted tool calls against the live environment — and only then does it create, update, narrow, delete or skip the record.
The tools it gets are a safe, read-only subset of the ones the task agent had: in the database setting it can inspect tables, join keys, encodings and post-migration fields; in the consulting worlds it can inspect file locations, document relevance, workbook contents and tool conventions. Step through the three phases.
The probe itself is not open-ended. The paper enumerates six questions the curator is asking, each of which is a question about the world rather than about the transcript. Toggle through them.
On the forty-question database schedule — which includes a deliberate schema migration partway through, so stale notes get punished — the four conditions separate on every axis at once. Pick a metric.
Two things in that comparison matter more than the headline. The first is that pasting in every prior transcript — the brute-force option — already gets you 61% and the cheapest query count, so a memory system has to beat that, not just beat nothing. Transcript-written memory beats it on pass rate and loses to it on reward; probed memory beats it on both.
The second is the spread. Transcript-written memory scored 70% with a standard deviation of 16 points; probed memory scored 73% with a spread of 5. The gain in the mean is three points. The gain in predictability is most of the variance, and on a production platform that is the number you would feel.
The database result is one environment. The second evaluation is ninety adapted management-consulting tasks spread over six independent worlds, scored as reward gained per dollar the task agent spent. Every one of the eighteen memory-versus-no-memory comparisons came out positive, and probing gave the best return in five of the six worlds. Step through them.
The mechanism behind that return is not that the agent gets smarter. It is that it stops rediscovering the environment: across the worlds, task-agent tool calls fall by between 16% and 75%. In the easiest world the fall is from 71.6 calls per run to about eighteen, and the run cost from $54.30 to single digits. Fewer calls is not only cheaper — on long-horizon tasks it is also what keeps the run inside its context and its step budget, so the reduction can buy correctness rather than just latency.
It also holds across model families, and the weaker the model, the more the probe is worth. On the thirty-question schedule without a migration, probing lifted mean reward more on the mid-tier model than on the frontier one — consistent with the paper's reading that probing pays exactly when the trajectory left a join, a file location or a procedure unresolved.
Take the paper's measured per-question numbers and run them against your own volume. Choose what you ship today, set how many long-horizon sessions a day your agent serves, and read off tasks passed, queries issued and task-agent spend. The per-task figures are the paper's; the volume is yours.
The shape worth noticing is that the cheapest option on paper — no memory — is the most expensive to run, because an agent with no notes rediscovers the schema from scratch every session. The probe does not add cost at task time at all. It spends read calls in a background process you already run, after the user has their answer.