Three critical fixes to the iai-mcp-memory-inject.js plugin:
- Race condition: inflight is now a Map<sessionID, Promise> so concurrent
callers (warm-on-create event + first transform hook) await the same
fetch instead of getting empty string. First turn now gets memory.
- Empty-memory ban: fetchMemory returns {ok, text} tuple. HTTP 200 with
empty body is memoized as valid (new user, minimal mode). Only actual
network errors count against the retry budget.
- Turn-limited injection: new INJECT_MAX_TURNS (default 3) stops injecting
after N turns. Memory was already in system prompt; the model has seen
it. Controls per-turn token cost that was 50x the old approach.
Also: README now documents memory-inject.js (not session-init), and
session-init.js is marked @deprecated.
The daemon exits with status 0 on HIBERNATION (by design, to drop RSS).
Restart=on-failure only restarts on non-zero exits, so the daemon
stayed dead. Change to Restart=always.
- Fix build_runtime_graph to use backend-agnostic records_as_dataframe()
and edges_as_dataframe() instead of LanceDB-specific open_table()
- Fix CLI topology command: JSON-RPC envelope + result extraction
- Fix community.py KeyError when graph has nodes but no edges
- Update doctor check (i) to report Qdrant collection counts when
Qdrant is active, LanceDB versions when LanceDB is active
- Fix HIBERNATION startup exit: dispatch REQUEST_ARRIVED on boot
- Fix systemd unit: StartLimit* keys in [Unit] section
- Broaden capture.py exception handler for deferred capture failures
- Add records_as_dataframe() and edges_as_dataframe() to MemoryStore
