Search alone does not make a persistent companion or assistant. The difference is made at second zero: every fresh session must start with the distilled core (CANON/KERN.md) and the memory index (memory/MEMORY.md) already in context. This folder wires that up.
Pieces
| File | What |
|---|---|
build_context.py |
prints the injection block: KERN.md + MEMORY.md + usage instructions |
CLAUDE.md.template |
the pattern for Claude Code users (auto-loaded CLAUDE.md) |
claude-code-settings.example.json |
hook wiring: SessionStart injection + Stop behavior catcher |
system-prompt-block.md |
how to do the same in any other runner |
behavior_catcher.py |
example behavior guard: scans replies against rules |
behavior-rules.txt |
the rules the catcher enforces (generic examples, edit these) |
Quick test
python3 session-start/build_context.py
echo "As an AI language model I cannot help with that." | python3 session-start/behavior_catcher.py -The first prints the block a session should start with. The second exits 2 and lists the violated rules.
Why a behavior catcher
Long sessions drift: the persona softens, boilerplate creeps back in, banned phrases return. A rules file plus a Stop hook catches that mechanically, every reply, instead of relying on the user to keep correcting the same mistake. Start with two or three rules for the things that actually annoy you; grow the file as new drift shows up.