Verify this release: https://d-o-hub.github.io/rust-self-learning-memory/verify-a-release.html
Release Notes
Fixed
- Turso monitoring metrics could never be written (#1180):
store_agent_metrics
and store_task_metrics named columns that the DDL does not create
(agent_metrics.last_execution_time, task_metrics.avg_completion_time,
vs. last_execution / avg_completion_time_ms), so every insert failed with
table … has no column named … on any freshly initialized database. The
reads used the same wrong names, so the failure was invisible: the only tests
covering the path are nightly-only and asserted result.is_ok() without
printing the error. avg_duration is now stored and read in the column's
declared unit (previously written as seconds into avg_duration_ms and read
back divided by 1000).
- Embeddings written under the
turso_multi_dimension schema were unreadable:
_store_embedding_internal, _delete_embedding_internal and
list_embedding_ids_backend are dimension-aware under that feature, but
_get_embedding_internal and has_vector_embeddings still queried the legacy
embeddings table, which the feature never creates, so every read failed with
no such table: embeddings. Reads now take the same dimension-aware path.
- Pooled Turso callers no longer use the direct-connection accessor: the
capacity-eviction intent table, the sync watermark and the embedding listing
called get_connection_with_id(), which returns
get_connection() is only available in direct-connection mode whenever a pool
is configured; they now checkout through with_connection().
- Nightly can no longer misreport feature-gated schemas as storage bugs:
multi_dimension_routing and phase1_validation are built only with
turso_multi_dimension (required-features), and nightly-tests.yml runs
those two suites in a dedicated step with the feature. Ignored per-test
cfg_attr(…, ignore) guards were not enough because nightly forces ignored
tests to run with --run-ignored only.
Tests
#[ignore]d Turso fixtures now match the default schema: the resilient
keep-alive tests stored 5- and 2-dimensional embeddings into the
F32_BLOB(384) column, and the metrics assertions printed a fixed sentence
instead of the real Err, which is why three production defects stayed hidden
while nightly was red.
Performance
- Turso-to-redb synchronization writes in batches (#1171): each keyset page is
chunked by SyncConfig::batch_size (default 100) and cached through one
RedbStorage::store_episodes_batch write transaction instead of one per
episode, so syncing 1000 episodes measured 205 ms against 1.66 s at
batch_size = 1 (8.1x) using 10 write transactions instead of 1000. A chunk
whose batch fails or exceeds SYNC_EPISODE_TIMEOUT falls back to per-item
stores for that chunk only, so one bad episode cannot mark unrelated episodes
as synced. The revision sequence stays allocated per episode: the cache
derives modified_at from it, so one shared revision per chunk would collapse
the whole batch onto one timestamp and let the keyset read fall back to its
episode_id tie-break, losing write order.
CI/CD
- Publishing to crates.io actually publishes again. The
publish-crate
composite guarded its authenticated cargo publish with !inputs.dry-run,
and because composite inputs arrive as strings GitHub coerces "false" to
true, so the live upload step was skipped on every tag publish while the job
reported success: releases v0.1.44 and v0.1.45 never reached crates.io
(do-memory-core was last published at 0.1.34, do-memory-mcp at 0.1.31).
The guard now compares against 'true' like the tag and auth steps, every
real publish is followed by a bounded check that the version is visible on
crates.io, the version-exists gate tests membership of the full version list
instead of equality with the newest entry, and the gate-contract and
publish-fixture validators fail if any of those three regress (regression of
#1109).
do-memory-cli 0.1.46
Download do-memory-cli 0.1.46
do-memory-examples 0.1.46
Download do-memory-examples 0.1.46
do-memory-mcp 0.1.46
Download do-memory-mcp 0.1.46