doobidoo/mcp-memory-service v12.3.0

on GitHub

Special Thanks

Thanks to @andrunio, @dodoshell, @ozenyunus35, @fatherlinux, and @guamiao for their contributions, and to @3099404236 for reporting the bootstrap authorization issue.

Highlights

• Bootstrap reads preserve agent session state; the bootstrap tool requires write scope (GHSA-xhjp-vf97-77wf).
• Codex reports memory load and save activity through system messages.
• OpenCode inbox polling catches messages across polling boundaries without repeating delivered rows.
• Milvus memory searches work again, and hybrid health checks expose Cloudflare sync failures.

Added

• Codex lifecycle hooks report memory load and save status through systemMessage, making successful hook activity visible in the session (#1648, dodoshell).

Fixed

• Bootstrap reads preserve agent session state (GHSA-xhjp-vf97-77wf).
  get_bootstrap_profile now requires write scope because it resets session
  counters when a fresh start is due. Read-only clients can use
  memory://agent/{agent_id}/bootstrap, which preserves the counter, timestamps,
  sync events, and fresh-start signal. Counter lookup no longer writes access
  metadata through semantic retrieval. Reported by @3099404236.
• Dashboard toggle, sliders and checkboxes still used the old blue (#1548, ozenyunus35).
  The checked toggle switch was hard-coded to #2563eb, and the range sliders and
  checkboxes had no accent-color, so the browser drew them in its default blue.
  body now sets accent-color: var(--primary), and the remaining #2563eb,
  #3b82f6, #60a5fa and rgba(59, 130, 246, …) values in style.css use the
  brand variables or rgba(25, 194, 218, …).
• OpenCode inbox delivery across polling boundaries (#1636, guamiao). The inbox re-queries a configurable overlap window and drops previously delivered rows, so messages that arrive around a polling boundary are picked up without being shown twice.
• The hybrid health check said syncing while every Cloudflare call failed (#1638).
  It reported cloudflare_sync: syncing whenever the background sync loop was alive, so a
  container with broken DNS synced nothing for 16 days without the health check noticing.
  It now reports sync_failing when the sync service reports Cloudflare unreachable, and
  exposes cloudflare_available and consecutive_failures next to the status.
• memory_search failed on every call with the Milvus backend (#1650).
  Since #1297 the MCP search handler passes agent_id to storage.search_memories(), None
  included, and Milvus is the only backend that overrides that method — its signature stopped
  at store, so each call raised unexpected keyword argument 'agent_id'. Milvus now accepts
  the keyword: without an agent the search runs as before, and a non-None agent_id returns an
  error result, as get_all_memories and count_all_memories already refuse it. The signature
  conformance test now checks agent_id on every backend, alongside store.

Internal

• PR CI runs shorter (#1642). A push to a PR now cancels the run for the commit it
  replaces; pushes to main keep one run per commit. The two Milvus legs run only when a
  PR reaches code the Milvus suite imports (by package prefix), and always on main.
  Each pip-installing job restores its own download cache, keyed on pyproject.toml.
• CI installs dependencies with uv instead of pip (#1643). The test jobs build a venv
  with a pinned uv from the runner's python3 and put it first on PATH; installs took
  18-77 s per job with pip. The per-job download cache now holds ~/.cache/uv.
• Tests share the default ONNX InferenceSession across a run (#1644). The autouse
  fixture that empties the embedding model cache after every test made each storage
  setup create a new ONNX session for all-MiniLM-L6-v2, about 0.3 s each. A
  session-scoped fixture now hands out one session per provider list for that model
  file only; everything else still reaches the real constructor. Locally the ML extras
  suite went from 537 s to 216 s and the default suite from 686 s to 228 s.
• The ML extras job prints its 25 slowest tests and setups (#1645). It is the
  longest CI job; --durations=25 puts the data for the next speed-up in every run's log.
• CI tests the default job against uv.lock (#1647). No job read the lock, so a
  Dependabot uv bump ran tests on exactly what main installed. Tests + Coverage and the
  prove-fix job now install with uv sync --frozen; the ML extras and Milvus jobs keep
  resolving from pyproject.toml to catch upstream breakage.
• Tests + Coverage and ML extras run in parallel with pytest-xdist (#1649). Both
  jobs use -n auto, one worker per runner core; each worker gets its own temp test
  database from tests/conftest.py.