The flattened text output now uses markdown headers so the model can
distinguish L0 identity from L1 summary from L2 episodes from handles.
This replaces undifferentiated newline-separated text with structured
sections: ## L0 Identity, ## L1 Recent Summary, ## Rich Club,
## L2 Episode, ## Handles.
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.
- _TableShim.add(): route by table name to dedicated upsert methods for all 5 tables
- _TableShim.to_pandas(): full DataFrame construction for edges/events/budget/ratelimit
- _parse_where: full parser with >, >=, <, <= (→Range), = (→MatchValue), AND splitting
- records_as_dataframe: include provenance_json, created_at, schema_version columns
- iter_records: wire columns parameter to Qdrant with_payload
- iter_record_columns: stop forcing 'id' into every yielded row
- events_query: scroll all events + Python-side since filter (handles missing ts_epoch)
- events_add: store ts_epoch alongside ts for numeric range filtering
- Module docstring: document _TableShim interface
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.
Cleans up duplicate edge points left by the boost_edges() bug (fixed
in previous commit). Each unique (src, dst, edge_type) keeps the
highest-weight point; duplicates are deleted. Default --dry-run.
Duplicate edges root cause — critical logic bug in QdrantStore:
boost_edges() and add_contradicts_edge() always created new points with
random UUIDs instead of reusing existing point IDs. Each boost call
added a duplicate point rather than updating in-place, causing 190
duplicate edge rows for 73 unique edges.
Sigma fix — compute_sigma() checked total node count (311 >= 200 floor)
but the largest connected component had only 2 nodes. Now checks the
largest CC size, which is the correct semantic for Humphries-Gurney
sigma validity.
The Qdrant _TableShim returned updated_at as ISO strings (not datetime
objects), causing str.replace(tzinfo=...) to fail with
"str.replace() takes no keyword arguments". Parse strings via
datetime.fromisoformat() before timezone handling.
Also add missing delete(where) and update(where=, values=) methods to
_TableShim so the LanceDB-compatible shim works with Qdrant's payload
collection, plus a _parse_where() helper to convert where clauses to
Qdrant Filters.
On non-darwin platforms, the wrapper now attempts to spawn the daemon
via 'iai-mcp daemon start' (which uses systemctl --user start) when
the daemon socket is unreachable. This allows the daemon to wake from
hibernation automatically instead of only writing wake.signal as a
passive marker.
If spawn fails (binary not found, systemd unavailable), falls back to
writing wake.signal for future daemon boots.
- 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
- doctor: skip LanceDB check when qdrant_storage/ dir detected
- topology: use daemon socket instead of local store (avoids lancedb crash)
- qdrant_store: add records_as_dataframe() + wire into _TableShim
so build_runtime_graph() works with Qdrant (was returning empty)
Move Quick start above Usage/How it works for faster onboarding.
Move About the name to end. Add license, platform, and benchmark
badges. Add table of contents for navigation.
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: XNLLLLH <XNLLLLH@users.noreply.github.com>
