01
Starting a new session
An agent starting a new session may have no access to the previous one. A teammate may have missed the meeting where a decision changed. I want both to be able to recover the current state without reading the entire chat history.
Once you accept that every member of the team has unreliable memory, the design question changes: not "how do we remember?" but "where does state live so that the next session — anyone's next session — starts from it?"
02
The living handoff
The practice: a context file lives in the repository and is maintained like code. Decisions, open threads, current experiment state, and the reasoning behind reversals land there — compiled from meetings, chat, and sessions rather than dying in the app where they happened. Agent-facing context files sit beside it, so a coding agent launched cold gets the same orientation a new human collaborator would.
The file is a contract with the future: whoever picks the project up next — including you, tomorrow — starts from the same state, at the cost of a few minutes of writing at the boundary of each session.
03
Context hygiene
Long-running conversations rot: they accumulate stale assumptions and half-superseded plans. So sessions get refreshed deliberately, and the handoff document carries state across the boundary. The goal is that no single conversation is load-bearing. If a thread dies — a chat resets, a teammate leaves, an agent's context fills — the project state survives, because it never lived in the thread.
04
A connection to memory research
Recall Debt asks about retrieval that depends on an intermediate bridge. A handoff file serves a similar purpose in my workflow: it connects the previous session’s evidence to the next session’s starting point. That is an analogy to team practice, not a result measured in the paper.
A handoff can also preserve a mistake. I want it to point to the underlying evidence and mark what is uncertain, rather than becoming an authoritative summary nobody checks.