feat: v9 brain modes — session focus, Ebbinghaus decay, dream consolidation, mid-turn recall

[cbuntingde]

Summary

Turns kimi-memory from a recallable filing cabinet into something closer to a working brain. All upgrades are wired into the hook layer; no new MCP tool is required for any of them. Adds a single new MCP tool (memory_reset_project) for the re-clone wipe case, and a new PostToolUse hook for mid-turn recall.

Version bump 0.4.0 → 0.5.0. Schema SCHEMA_VERSION = 9.

Brain mode (the four behavioural upgrades)

• Continuous retrieval — every UserPromptSubmit recall now unions prompt tokens, working-memory slots, the session-focus title, and recent file paths from conversation_events. The top 3 hits are round-robined across memory types so a single type doesn't crowd out the rest.
• Mid-turn recall — new PostToolUse hook surfaces up to 2 stored conventions matching the tool's arguments (file path stems, shell verbs). Cheap — no LLM call. Degrades silently on Kimi versions that don't declare the event.
• Ebbinghaus decay — every memory carries a per-row stability_days and last_rehearsed_at. Confidence is rewritten on SessionStart from 0.1 + 0.9 * exp(-days / stability). memory_reinforce grows stability by 1.5× (cap 365) and stamps a fresh rehearsal. The hook auto-reinforces the top project recall hit with a 60-second debounce so the feedback loop is closed without manual calls.
• Cross-session narrative — SessionStart lists the last 3 sessions for the project (oldest → newest) with each session's focus title and body snippet.

Session focus — "where we left off"

• Stop hook writes a working-typed memory titled Last focus: <truncated latest user prompt> whenever the session has at least one user prompt. Tags: focus, session-focus, in-flight. expires_at = now + 30 days. Zero LLM cost — built from conversation_events.summary. Disable via KIMI_MEMORY_DISABLE_SESSION_FOCUS=1.
• SessionStart and UserPromptSubmit surface a [focus] line so the next session has the thread to pick up.

Background consolidation ("dream pass")

Every SessionStart clusters active memories by embedding cosine (≥ 0.75) and tag overlap (≥ 2 shared tags), and for each cluster of ≥ 3 siblings with no existing conclusion child writes one synthesised conclusion. Idempotent via memory_synthesizes coverage check. Opt out with KIMI_MEMORY_CONSOLIDATE=off.

memory_reset_project (re-clone wipe)

memory_reset_project(cwd, confirm?) wipes every per-project row (memories, working memory, conversations, conversation events, edges, synthesizes). The project_key is a hash of the canonical path, so a re-clone is otherwise indistinguishable from the original. confirm: false is a dry run. Global DB and other project DBs are never touched. The hook layer surfaces a [stale-memory] line on SessionStart and UserPromptSubmit when directory birthtime is newer than first_seen_at. CLI gets a matching reset-project subcommand.

Persist refactor

Collapses src/persist/{conversations,db,edges,memory,working}.js into a single src/persist.js. The five modules were a single SQLite schema split across five files; the new layout keeps the table openers and migration helpers in one place. Public API (saveMemory, listMemories, …) is unchanged. Net −1028 lines across the persist module.

Schema

SCHEMA_VERSION = 9. Migrations add stability_days and last_rehearsed_at to memories. Existing rows backfill on first open (stability = 30, last_rehearsed_at = updated_at).

Tests

3 new test files:

• tests/22-brain-modes.test.js — continuous retrieval union, diversification round-robin, Ebbinghaus decay curve, tool-recall path/verb triggers.
• tests/22-reset-project.test.js — detectReclone branches, resetProject isolation, MCP round-trip dry-run + apply, end-to-end re-clone wipe.
• tests/23-session-focus.test.js — capture, dedupe within session, expires_at enforcement, focus line shape, integration with SessionStart / UserPromptSubmit.

247 pass, 1 skipped, 0 fail.

Why this approach fits

The plugin already had the primitives — conclusion typing, memory_synthesizes, memory_reinforce, memory_edges — but none of them were creating themselves automatically. The same goes for the [focus] / [recall: i/N] surface area: the hook layer already had the plumbing, it just wasn't pulling from enough cues. v9 wires all of those into the hook lifecycle and lets the user opt out per-knob via env var, matching the KIMI_MEMORY_AUTO_EXTRACT / KIMI_MEMORY_SECRET_SCAN pattern.

The persist refactor is a precondition for several of these — schema v9 needs ALTER TABLE memories ADD COLUMN stability_days to live alongside the other migrations, and the previous five-file split made it awkward to keep the table openers in sync.

Checklist

• Tests added/updated — three new test files, all green
• package.json version bumped (0.4.0 → 0.5.0)
• kimi.plugin.json version + longDescription updated
• README documents new env vars, new tool, and the brain-mode section
• skills/kimi-memory/SKILL.md updated
• No secrets, no .env touched
• No new top-level dependencies
• CI green (watching after opening)