Repository navigation
v2.22.0
An audit of two weeks of real use found the automatic memory loop mostly idle: no memory had an embedding, prompt-time injection did all retrieval but never counted as use, every session-end proposal was a raw tool error, and auto-promoted memories waited for a review nobody did. This release makes those loops work without the user or the agent having to remember to run them.
Changed
- Session-end proposals come only from the LLM pass. With
ltm.evaluateSessionLlmoff (the default), raw tool errors stay inlearned/patternsand nothing is queued for review. In real use, 13 of 13 proposals built from tool errors were noise (fatal: not a git repository, test output, hook errors). - Auto-promoted memories are accepted after 7 days. The janitor gives decisions and gotchas promoted from project context a title and origin, then approves any still pending after a week (audited as
janitor:auto-accept). SessionStart lists the ones accepted this week, withforget <id>as the undo. - Prompt-time injection counts as use. Memories the UserPromptSubmit hook injects now update
last_used_at,last_recalled_atandrecall_count, the same as arecall. SessionStart updates only the memories it actually injects;getContextMergepicks the project's top memories by decay score instead of 15 arbitrary rows. - Project names are case-insensitive. Every write lowercases the project scope, and migration 031 folds stored casings (
SummonOSandsummonoswere two projects). A git repository inside a registered folder now keeps its own name instead of taking the folder's (~/Documents/OpenLtmresolved todocuments). - Stale flags no longer touch preferences or owner statements, and flags already set on them are cleared. A memory flagged stale by a commit and not revalidated within 30 days is deprecated (importance 5 exempt).
src/schema.sqland the rootmigrations/copy are gone;packages/openltm-coreholds the only copy.getSimilarMemoriesindb.tsis renamedgetSimilarMemoriesByVec.
Fixed
- Embeddings never drained. With Honker available,
learnqueues the embedding and only the graph server ran the worker, so a plugin without the visualizer never embedded anything and recall was full-text only. The MCP server now runs the worker too. SessionStart says when semantic recall is off because no embedding provider answers. - Staged conflicts were never detected. The janitor now runs contradiction detection and stages pairs for review, so SessionStart's conflict section has data.
- The git post-commit hook could not open the database. The bundle inlined core, which then looked for
schema.sqlnext tohooks/. Core is now external to the bundle. - A database from an older release (e.g. 1.8.0) failed to open:
schema.sqlindexes columns a later migration adds. It is now migrated first. - Processes opening a brand-new database at once raced on migrations (
table _honker_notifications already exists). Migrations run under a<db>.migrate.lock. SQLITE_BUSY_RECOVERYwhen a process opened the database while another wrote.busy_timeoutis now set beforejournal_mode=WAL, and the graph server sets one at all.- OpenCode on macOS: the installer created
~/Library/Application Support/opencode/opencode.json, which OpenCode does not read. It now uses~/.config/opencode. configure()with a new database path dropped nothing, so the old handle kept serving the old database.- The MCP server exits when its stdin closes.
Tests & CI
- CI runs typecheck and tests on macOS as well as Linux, and
qa:smokeon Linux. The suite passes on macOS (10 tests assumed Linux temp paths or a case-sensitive disk). - New tests: hooks with a corrupt database exit 0, concurrent writers and concurrent first open, upgrade from the 1.8.0 schema, the Hermes schema against core's, and several assertions that passed on
null.