When we first ran an agent across multiple sessions, the instinct was to save almost everything. The agent had produced something. Discarding it felt wasteful. A year later, we throw away most of what an agent generates within minutes of generating it. The rest gets quietly retired before the next session begins. What survives long enough to matter to a future run is a much smaller set than we expected.
The change came from a slow accumulation of failures, not a single insight. An agent that “remembered everything” got worse over time. Its retrieval started returning stale assumptions alongside current facts. Decisions made under one set of constraints came back to bias decisions made under another. The store filled up, the embeddings drifted, and the agent started disagreeing with itself across sessions in ways we couldn’t easily debug.
We now think about agent memory in three rough tiers, less because the taxonomy is elegant than because it gives us three different rules for what to keep. Each tier answers a different question about the cost of forgetting versus the cost of carrying a stale fact forward.
What lives inside a turn
The first tier is the scratchpad of a single turn. Raw tool outputs, intermediate reasoning, the half-formed plan the agent considered before settling on the one it executed. We keep all of it while the turn is running, because the agent might still need it. We throw almost all of it away the moment the turn ends.
The exception is whatever the agent has committed to in the final output. If the agent ran a search and decided which source to cite, the citation survives. The other four search results do not. If the agent ran three different SQL queries to figure out the shape of a table, the answer survives. The queries do not, unless they were the actual deliverable.
This sounded obvious in principle and was hard in practice. An agent that re-reads its own scratchpad on the next turn behaves differently from one that doesn’t. Tool outputs from the previous turn are recall-shaped, not action-shaped. They look like information the agent might want again, so the agent uses them. They are almost always slightly out of date. We learned to prune more aggressively at the turn boundary than we initially wanted to. It costs one re-fetch in the rare case the agent actually needs the value again. The cost of carrying it forward is paying for stale information to compete with fresh information on every subsequent retrieval.
What survives a session but not the next
The second tier is the working set inside a session. Plans the agent made, hypotheses it formed, branches of reasoning it explored and then abandoned. These don’t make it past the session because most of them were wrong on the way to being right.
A retracted hypothesis is the clearest case. Halfway through a session, the agent thinks the problem is in component A. Two turns later, it has evidence the problem is in component B, and component A is fine. If we promote the early hypothesis to long-term memory, the agent will keep finding “evidence” that component A is suspect for weeks after the original session closed. Embeddings don’t care that we changed our minds.
We’re stricter at this boundary than we are inside a turn, because the cost of getting it wrong is higher. Inside a turn, a stale fact lives for seconds. Inside the long-term store, it can live for months and quietly bias every related decision. So the rule at the end of a session is closer to “nothing survives unless we can name why it should.” Plans don’t survive. Hypotheses don’t survive. The final outcome of a chain of reasoning sometimes survives, if we can describe it as a fact rather than as a story. “The user prefers terse responses” is a fact we can keep. “The user seemed frustrated when we wrote a long response on Tuesday” is a story we shouldn’t.
What we are willing to write down for a long time
The third tier is what we promote to long-term semantic memory: stable preferences, durable constraints, lessons that have survived being contradicted. The bar for getting in is high. The set is small. It’s small on purpose.
A preference the user states once is not eligible. A preference the user has stated in two different sessions, in two different framings, with no contradictory evidence in between, is. A constraint that came from a specific deadline last quarter is not eligible after the deadline passes. A constraint that came from a hard limit of the system we’re integrating with is.
The discipline that matters here is less about what to add than about what to evict. We try to make every long-term fact answerable at retrieval time: is this still true. If it isn’t, the fact has to go. We’ve found that the cost of letting a falsified fact survive is high enough that we’d rather lose a true fact than keep a false one. Memory drift, in our experience, is mostly the slow accumulation of facts that used to be true.
We do not have a fully automated answer to eviction. We have rules of thumb. If a long-term fact has been contradicted by direct evidence in a session, we mark it for review. If a fact hasn’t been retrieved in any session for a long interval, we mark it for review. If a fact’s score on a semantic match keeps creeping up while the agent’s behavior gets less appropriate, we mark it for review. The review itself is sometimes a question the agent asks Igor and sometimes a question another agent answers from session transcripts. Nothing leaves long-term memory silently.
What the discipline keeps protecting against
The thing we keep wanting to do, and keep deciding not to do, is “remember more, just in case.” It feels safe. It costs almost nothing in storage. It is the failure mode underneath almost every memory problem we’ve seen.
The store is not the cost. Retrieval quality against the store is the cost. Every additional fact that lives in long-term memory competes with every other fact for the same attention budget at the moment of retrieval. Doubling the number of facts roughly halves the signal-to-noise ratio of the average match. We notice the loss as the agent quietly getting less specific over time, not as a single moment of failure.
So we keep less. We treat working memory as exhaust. We treat session memory as a draft we don’t save. We treat long-term memory as a small reference book we are willing to defend the contents of, fact by fact. The agents that work best for us are not the ones with the most memory. They’re the ones whose memory is mostly empty, on purpose.