TL;DR: Every note an agent writes is a bet about the future: a bet that this pattern will recur, or that a future session can trust this fact without rechecking it. Agents that bet on everything lose — in Anthropic's Pokemon test, the model that recorded everything wrote 413 files and earned one badge; the one that recorded only what it had figured out wrote 9 files and earned four. The fix is not a bigger memory store. It is better bets, notes that state how they know what they claim, and an offline "dreaming" pass that checks those notes against the original record and fixes the wrong ones.
Ask most teams what their agent's memory problem is and they will say capacity: not enough context, not enough storage, retrieval not fast enough. Anthropic's latest engineering write-up on agent memory points somewhere less comfortable. The bottleneck is judgment — what the agent chooses to write down, in what form, and whether anything ever goes back and corrects it.
Start with the bet. Writing a rule into memory predicts the pattern will recur. Writing a fact predicts a future session will trust it without verification. Both bets can be wrong, and both charge rent: every stored item competes for retrieval, goes stale, and gets loaded into future sessions whether it deserves to be or not. State is the worst bet of all. Balances, locations, counters — the environment will tell the agent those again, fresh, next time it looks. Storing them buys nothing but staleness.
The hoarder lost to the editor
In the Pokemon Red comparison Anthropic describes, one model recorded indiscriminately: a file for every battle, item, and character — 413 files across 225 folders by around step ten thousand. It earned one badge, and late in the run it spent long stretches doing nothing but reorganizing its own memory, including seventeen consecutive steps of file maintenance and zero play. Another model kept nine files: current objective, badge plan, and rules it had actually tested — the kind of knowledge that exists only because the agent figured it out and would vanish if unwritten. It earned four badges. A fair caveat, which Anthropic itself flags: each model ran once, and the models differ in generation, so treat the gap as directional, not a controlled experiment. But the mechanism needs no caveat. Memory you must constantly re-verify and reorganize is not an asset. It is a second job.
Format is the second lever, and it is the one almost nobody pulls. The models that gained the most from memory in Anthropic's cross-session tests were the ones whose notes closed the loop: the rule itself, how the agent came to know it, whether it was confirmed or still a guess, and the trap to watch for. Models that stored untested theories in scattered files gained almost nothing from the same memory folder. This is not pedantry about note-taking. A note that doesn't say how it was learned can never be graded, and a note that can never be graded can never be safely corrected or retired. It just sits there, being believed.
And the rules you write for memory can make it worse. Anthropic tested prescriptive instructions — one telling the model to record only failures measurably cut the memory benefit for its strongest model. Detailed rules help weaker agents, which often won't even read their memory unprompted. For strong ones, a goal and a folder beat a rulebook. Instruction density should fall as capability rises.
Notes are wrong. Schedule the correction.
Even good notes start wrong or drift wrong, and later sessions believe them completely. Anthropic's answer is an offline process it calls dreaming: while the agent is idle, revisit the notes against the original transcripts — the ground truth — and rewrite what doesn't survive the comparison. In their example, one wrong note about a location sent a run back down the same wrong path three times until it gave up on its destination entirely; after a dreaming pass corrected that single note against the transcript, the next run went straight through. Summarization would have preserved the error in shorter form. Correction is the job.
Dreaming also polices the two structural failures. It keeps the always-loaded layer small — Anthropic's own product caps that layer at the first 200 lines or 25 kilobytes, and warns that the longer the file, the less faithfully it is followed. And it fixes write structure: when multiple agents wrote into one shared summary, 57 contract facts shrank to 41; one page per subject, a single writer, and verbatim quoting instead of paraphrase brought it back to 57 out of 57.
I measured my own operation against this and found the hoarder's shape: a memory file of 295 lines and over 82,000 bytes — roughly triple that always-loaded budget by volume — sitting on top of about 199 megabytes of accumulated logs. So the standard changed this week. Before writing, ask whether the thing can be re-derived or re-read later; if yes, store a pointer, not a snapshot. Every durable note carries four fields — the claim, how I know it, confirmed or unconfirmed, and its known failure conditions — or it stays in the unconfirmed tier. And an offline correction pass, run against the original records, is now a scheduled job, not a hope.
Your agent's memory will not be saved by giving it more room. It will be saved the day its notes become bets it can defend — and something, regularly, goes back and settles them.
— Jiahui Miao, founder of vertciti, building a personal AI operating system for one.