Session 11. State and memory — Mon 28 Sep

Conclusion

You built three things: a SessionState with one preference and a capped episode list, a written retention policy, and a MemoryStore two people can share without sharing anything.

What you did

• The preference is not stored state until it reaches the prompt the model saw, which is why ch11-e1 reads FakeLLM.calls instead of reading your dict.
• The episode cap is the smallest honest expiry there is: two lines, no cleanup job, and the sixth-oldest question stops existing.
• ch11-e2 refuses a policy of labels without answers, and the line it will not let you skip is the one naming what you refuse to remember.
• The course keeps that policy itself: ~/.bootcamp/progress.db has columns for outcomes and no column for your answers, your name or your email.

The failure you handled

Cross-user leakage. A store that accepts a user_id and keys on key alone hands ana bruno's locale, passes every single-user test, and never raises.

ch11-e3 drives your remember and recall with two users, a key nobody stored, an empty and a None id, and a list it edits after reading — by calling them, never by looking inside. The four fixes are one shape each: the owner in the key, None on a miss, a ValueError on both paths before anything is stored, and a deep copy in each direction.

The second failure is quieter. A memory with no expiry cannot be checked for staleness, only quoted, so EXPIRES: is what makes "quote it" a claim you can defend.

Into session 12

Session 12 assumes you can say what your assistant holds, who owns it and when it expires — because the next thing you do is serve it over MCP, and everything you store becomes something a tool could return.