Memory

The assistant retains nothing between sessions. Everything that behaves like memory is built outside the model, in four stores it reads and writes. What a session actually loads from them is a design decision of its own.

Choosing what loads

Context engineering is that decision, made deliberately. Load too little and the session re-derives what the practice already knew. Load too much and the load competes with the work. Here a session starts with two things only: the standing instructions on how to work, and the state document left by the session before. Everything else is read when the work needs it.

The four stores

The knowledge base is the deliberate memory: what was worth keeping, in checkable form, graded as the trust page describes. It answers questions of substance.

The change history is the involuntary memory. Every change to every file is recorded with its reason, at the moment it happens. It answers what changed and why, months later, and the assistant can read it at any depth.

The state document is the working memory between sessions. At the close of a session it is rewritten: what was finished, what stands open, who owns each open item, and what comes next. The next session reads it before anything else and continues from it. Open items move through the chain by identity: each one carries a key, and it appears in every following state document until it is closed. Closing is its own act, recorded on a closed list. An item cannot fall out of the document by being forgotten. It travels forward under its key, or it leaves through the closed list.

The session log is the memory nobody curates. Every session is stored in full, each message and each action, as structured files. Daily work never reads them. The log is a research object: a complete record of how the work is actually done. It can answer how error rates move over time, where corrections concentrate, and how the collaboration changes when the system changes. Questions not yet asked can be answered later, because the material is complete and keeps accumulating.

If you are not using Git, you're doing it wrong.