Problem
The model cannot learn, so anything a session taught you evaporates when it ends. Extract Knowledge captures insights when you notice them — but noticing in the moment is unreliable, and nothing looks across sessions for what keeps coming back.
Pattern
Make capture a ritual:
- End every session with a learning pass: the agent reflects on the session — corrections you made, surprises, gotchas, things you had to repeat
- Categorize and route each learning to the right durable store:
- behavioral rules → ground rules
- project facts → knowledge and reference docs
- recurring quality corrections → habit hooks
- repeated processes → a skill or automation
- You approve the routing — the agent proposes, you decide
- Advanced: also append a dated entry to a session log. Periodically have the agent read the accumulated log and surface recurring issues that deserve promotion into stronger mechanisms
Every session then leaves behind a mechanism the next one runs on.
An open-source implementation is available at https://github.com/devill/ivetts-skills#learn
Example
A /learn command runs at session end and proposes: "you corrected mock-heavy tests twice → habit hook candidate; the deploy needs the VPN → project ground rules; the release process we walked through → new skill."
A /remember variant first appends the session to a log, then reflects across the whole log. After a few sessions it notices the same flaky-test complaint recurring, and suggests fixing the root cause instead of re-learning the workaround every session.
