What the machine keeps

There is no single memory here. There are five, they cover different spans of time, they were shaped by different amounts of human judgment, and each one answers a different kind of question. Most of the skill is knowing which one a question belongs to, because asking the wrong one returns an answer that looks fine.

Read the bars against the scale at the top. Two of the five end, and those two are the ones a session actively manages. The other three run the length of the practice and differ instead in the right-hand column, in how much human judgment shaped them. A question belongs to whichever memory covers its span and whose shaping matches the kind of answer wanted.

What each one can and cannot answer

memory	answers	cannot answer	what it costs
the context	what is in front of us right now, and what we just agreed	anything from before this session. It will produce something anyway	the load competes with the work, and it degrades as the session runs
the state document	what is open, who owns it, what happens next	why a thing was decided the way it was, beyond one line	a few minutes at every close, every time, including the closes that felt uneventful
the curated base	what is known and how far it can be trusted	what happened. It holds conclusions, not events	the judgment to keep it small, which is harder than the writing
the change record	what changed, when, and the reasoning at the time	what is true now. It holds every step, including the ones later reversed	a written reason at each change, when the reason is still in your head
the session log	how the work is really done, over months	anything, until someone asks it a question and builds the means to read it	storage, and the discipline not to treat it as working memory

The one that degrades

The context is a memory with a health, and it is the only one on the list that gets worse while you use it. Two properties matter in practice.

The first is that drift does not correct itself. Once a conversation has entered a confused or cluttered pattern, each side tends to continue in that pattern rather than restructure. Responses inherit the framing of what came before, and the person on the other end carries their own momentum in the same way. This is a documented property of multi-turn work with these models rather than a quirk of one setup. The consequence is that the opening exchanges of a session carry more weight than their length suggests, and that a session which has gone wrong will usually stay wrong.

The second is that context is summarised when it runs out of room, and summarising is lossy by design. Anything that must survive verbatim is written to a file before the limit is near, not left in the conversation to be compressed.

signs the session has agedpoints already settled get re-argued. The grip on what is load-bearing loosens. Confidence rises in things asserted from memory instead of re-read. The work drifts back to assumptions the session had already moved past.

reset beats correctioncorrecting a drifted session spends turns inside the drifted frame. Starting a new one and carrying forward only what matters is usually the cheaper move, and the state document is what makes it cheap.

the cost is not the same every timea session doing close work on loaded material can run a long time. A session pulling in bulk new material cannot. Budget by the kind of work, not by the clock.

The one you have to be strict about

The curated base is the only memory where more is not better, because everything in it has to be maintained, read past, and eventually re-checked. The test before writing anything into it is whether a capable model produces the same thing cold. If it does, storing it buys nothing and costs upkeep, and that bargain gets worse with every model release as the floor rises underneath it.

What survives that test is the material a model cannot supply: judgment, specifics that were audited against a source, and the reconciliations where a standard says one thing and practice holds another for a stated reason. Those do not commoditise. A base built out of them stays valuable as the models improve, and a base built out of retrievable facts is worth less every month.

The second rule is that a fact lives in one place. Copying it into three related documents feels like good cross-referencing and creates three things that will disagree after the first correction. Link instead, and let the correction land once.

The nearest memory answers anyway

Nobody routes a question to the wrong store on purpose. The failure is quieter than that. An assistant answers from whatever is already in its context, because that is the cheapest place to answer from, and the answer arrives in the same confident register as one that was actually looked up. Nothing in the reply distinguishes the two.

what it says it didasked whether something is finished, an assistant will answer from its own intent as readily as from the result. A report of an action is not evidence of the action, and the two are worded identically. Check the artefact, not the account.

what came out of the intake pilethe material waiting to be checked is often the assistant's own earlier output. Reading it back as settled is self-confirmation wearing the appearance of a source, and the work built on it looks like work built on the validated base.

what was said earlier in a long sessiononce a conversation has been summarised, the wording is gone and only the gist remains, while confidence in what was said does not fall to match. The move at that point is to re-read the file rather than ask the session.

The move that fixes all three is the same, and it costs one step. Ask for the source rather than the answer. Open the record and say what it says. Show the note and its grade. Read the file back before acting on it. The instruction changes which memory gets used, which is the thing you cannot tell from the reply.

All five exist because the assistant is not continuous and the practice is. The person is the only part of this that runs the whole length, which is why the stores are built to be read by someone arriving with nothing.