Long-term memory

Durable facts about the user or household, remembered across sessions — distinct from notes/reminders (checklist items) and from the in-session conversation thread (short-term, not a "feature" so much as core plumbing).

How to use it

"remember that I'm allergic to shellfish"

"what do you know about my allergies"

"do you remember my birthday"

How it works

• Tools: save_memory (fact, optional category: identity / preference / health / home / note) and recall_memory (query) in brain/tools.py.
• Storage: the Mnemosyne capsule (brain/longterm.py → python/mnemosyne/). Full design: memory-capsule-spec.md.
• Two write paths, one destination: tool-driven save_memory calls and a heuristic extractor (longterm.remember_from_utterance, run on *every* utterance regardless of whether a tool fired) both write through brain/longterm.py, so there's one source of truth for storage even though there are two triggers for it.
• Every reply's system prompt is seeded with longterm.recall(command, user_id) so the character "knows" relevant facts without an explicit recall tool call.
• Episodic (within-session) conversation history is a separate, unrelated system — see brain-memory.md.

Status & known limits

• Double-write risk: the heuristic extractor runs *alongside* the save_memory tool, so the same fact can occasionally be stored twice via two different triggers. Open TODO: let the tool own writes, retire the heuristic path (or vice versa).
• No delete/forget tool yet, despite forget that appearing in the tool-intent gate — that phrase currently just makes the gate fire, with no tool behind it.