Long-term memory chose its candidates before it looked at the question.
Recall listed 400 items by `importance DESC, valid_from DESC`, loaded the
vectors belonging to that list, and scored those; extraction did the same
with a 200-item pool. Importance is uncorrelated with relevance, so the
vector stage could only re-order what importance had already picked, and a
memory below the cut was unreachable no matter how well it matched. The
capacity cap goes up to 2000, so everything past the 400th most important
memory was permanently unrecallable, and `memory_item_embeddings.vector`
being a BYTEA blob meant pgvector was never involved.
The store now answers "which of this person's memories is this question
about":
- Migration 000095 (and the SQLite mirror) add `embedding halfvec` plus a
`(tenant_id, subject_id, model_id, dims)` index. The column is
conditional on the vector extension, so a deployment whose retrieval
driver excludes postgres keeps working. No HNSW: an approximate index is
filtered after the fact, and one subject holds at most a couple of
thousand rows, so exact ordering inside the scope is both correct and a
few milliseconds.
- `SearchItemsByVector` ranks in SQL over every vector the subject has,
with the status, expiry and kind filters joined in so reaching past the
candidate list cannot reach past the filters. SQLite and pgvector-less
PostgreSQL fall back to scoring in process, still over the whole scope.
- Recall fuses the lexical ranking with those hits and grows the candidate
pool with matches the pool did not hold, which is what lets the search
escape it. The lexical pool is now the capacity cap, so it can no longer
be a second, invisible limit under the one that decides what stays active.
- Extraction picks the memories it shows the model the same way. This
matters more than recall: the memory a new statement contradicts is not
reliably an important one, so an importance-ordered window hid exactly
the collision that stops a duplicate being written.
- Vectors are written to the searchable column in the upsert's transaction;
rows predating the column are moved by `SyncVectorColumn` during
maintenance, which costs no model calls. The embedding backfill goes from
50 to 200 per run — at 50, a subject at the capacity cap took over a
month to become fully searchable.
- Traces and logs report `outside_pool`, the number of semantic matches the
lexical pool did not contain. It was structurally zero before, so it is
the number to watch when asking whether semantic recall does anything.
Tests cover whole-scope ranking, the lifecycle and kind filters, pool
growth, the pool-size invariant, and extraction seeing a low-importance
memory it would otherwise duplicate. A pgvector integration test
(WEKNORA_MEMORY_TEST_POSTGRES_DSN) exercises the migration, the halfvec
write, the SQL ranking and the column sync.
* feat(mcp): add per-tool enable toggle
* style: format migration test
* fix(mcp): keep enable and approval updates from overwriting each other
Partial policy writes now touch only the supplied columns, and a missing
tenant fails closed so a disabled tool cannot slip through.
---------
Co-authored-by: myysophia <myysophia@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
Accumulate token usage across every LLM round of an agent turn and
expose the aggregate everywhere a turn is observable:
- AgentState.TurnUsage accumulates each round's reported usage,
including the final-answer synthesis call; TokenUsage.Accumulate
keeps the cache counters as prompt subsets and recomputes the
combined cache status (unsupported survives only while every call
was itself unsupported).
- The agent completion event carries the aggregate, and the complete
stream event exposes it both as data.usage and as a typed
StreamEvent.Usage that buildStreamResponse promotes into the
previously-unused StreamResponse.Usage wire field.
- The assistant message persists it as a nullable usage column on
both migration targets (postgres 000085 jsonb, sqlite 000012 text),
for web and IM channels alike, so history reads still carry the
turn's cost after the live stream is gone.
- "[LLM Usage]" log lines gain ", session_id=..., principal=type:id"
when the context carries them; calls outside sessions stay
byte-identical.
Usage from failed retry attempts is not accumulated: the per-call
usage log remains the ground truth, while the turn aggregate serves
as the operational signal.