What it does

Captures knowledge at the end of an AI session so nothing is lost between contexts. A structured 7-step protocol that transfers what was learned, what was decided, and what remains unfinished into durable storage.

The pattern

When ending a session, run through these steps in order:

1. Review accomplishments. List what was completed. Be specific: "deployed feature X" not "worked on stuff." This becomes the commit log of your session.

2. Capture timeless patterns. Anything learned that will be true next week, next month, next year. API quirks, architectural constraints, failure modes. These go into persistent memory files.

3. Update persistent memory. Write new insights to long-term memory. If you discovered that an API silently truncates responses at 100 items, that fact needs to survive this session.

4. Flag unfinished work. For each incomplete item, write enough context that a fresh session (or a different person) can pick it up without re-discovering anything. Include: what was attempted, where it stalled, what the next concrete step is.

5. Note decisions and rationale. Decisions without recorded reasoning get revisited and relitigated. Write down what was decided and why. "Chose SQLite over Postgres because this is a single-user CLI tool with no concurrent access" prevents someone from re-opening that discussion.

6. Identify skill candidates. If you did something reusable during the session, note it as a potential skill or script to extract later. Do not build it now. Just flag it.

7. Write handoff summary. A clean 3-5 line summary: what was done, what is pending, what the next session should start with.

Each step has a destination: memory files for long-term knowledge, task systems for follow-ups, skill directories for reusable patterns, and the handoff summary for immediate continuity.

Key decisions

Protocol over habit. Running through a checklist catches things that "I'll remember that" misses. You will not remember it.

Separate timeless from temporal. "The deploy pipeline takes 4 minutes" is temporal and will change. "The deploy pipeline requires a clean git state" is timeless and belongs in memory.

Enough context to resume, not a novel. Unfinished work notes should let someone start working in under 2 minutes. If they need to re-read an entire codebase, the note failed.

When to use it

At the end of every non-trivial AI session. Skip it for quick lookups or single-command tasks. The protocol pays for itself the first time you open a new session and find a clear handoff instead of a blank slate.