CodeAbra/iai-personal-memory-engine v3.0.2

on GitHub

Added

• A hand-set behavioural-profile knob now persists and
  is pinned the instant it is set, rather than waiting for a later write to
  flush the whole profile. A pinned knob is recorded in the durable blob at
  set time and survives a process restart even if no other write happens
  first.
• The task-support knob now auto-tunes from genuine
  same-session follow-through: if adjacent suggestions go consistently
  unused it switches to blank recall, and if that later stops fitting, a
  bounded nightly probe (at most once every two weeks) briefly re-shows
  suggestions and switches back only when the user actually uses them
  again. An unused probe changes nothing.
• The monotropism-depth knob now auto-tunes per
  topic, keyed by the topic's own human-readable name, from how
  concentrated recall cues are on one topic versus a spread of topics
  night over night; a genuinely dominant topic's rank gain grows
  (bounded, capped well below the level a manual override would need to
  reach), and a topic that stops recurring fades back out. The knob's
  proactive same-topic duplicate check now resolves the same way, so a
  hand-set depth on a named topic reaches both the rank gain and the
  duplicate check consistently.
• A hand-set monotropism-depth above the manual
  threshold now also reorders which hits are served first, promoting
  memories from that named topic (including exact-match hits) to the
  front of the response. This reorder is manual-only; ordinary
  night-over-night auto-tuning never reaches it.

Removed

• Dropped the masking-detection tool and its behavioural
  knob. The camouflaging_status MCP tool is gone, leaving 14 tools on the MCP
  surface, and the camouflaging_relaxation knob is gone from the sealed
  behavioural profile registry, which now holds 10 entries. The recall path no
  longer applies a formality-relaxation adjustment based on that knob. Any host
  still calling camouflaging_status or reading/writing the removed knob must
  update to stop referencing them; historical audit-log rows recorded while the
  feature was active remain queryable. Thanks
  @tom-shields-stitch, whose analysis
  showed the knob sitting at seed defaults on a real 33k-record store and whose
  dead-path reports drove the surrounding cleanup.

Fixed

• A reader could briefly see fewer stored memories than were actually
  committed while writes were in flight — in the worst case a batch of recent
  memories vanished from one read and reappeared on the next. A read-only
  snapshot captured the write-ahead log in a single unguarded pass; racing a
  live writer could tear that pass mid-file, silently truncating the view to an
  older prefix that was then served as current. Snapshot capture now verifies it
  consumed the log to its true committed tail, a refresh can never regress a
  reader onto an older view, and a capture that cannot be verified fails loud and
  falls back to the strongly-consistent path instead of serving a phantom
  snapshot.
• Defense in depth for the same class: every cached row-count formation point now
  re-checks its freshness fence across the count read and recounts instead of
  caching when the fence moved, so a stale count can never be paired with a
  current generation.
• A stored memory whose internal vector-index label was
  ever left empty could not be recovered on the storage engine: the boot-time
  repair meant to backfill it silently did nothing (a column-to-column update
  wrote empty instead of copying the row key), so the store refused to reopen
  instead of self-healing. The update now copies the value correctly, so such a
  store recovers on the next start.
• A cached filtered row count could report more rows than
  actually match after a soft-delete, because the cache was not refreshed when a
  non-indexed column changed. The boot decision about whether to rebuild the
  vector index, the wake sequencing, and the corpus count all read that count, so
  a stale over-count could trigger needless work or mask a real mismatch; the
  count is now refreshed on every write that can change it.
• Setting a behavioural-profile knob to the value it
  already held recorded no pin, leaving it open to being moved by the
  nightly self-tuner even though the owner had explicitly locked it in.
  Any explicit set now pins the knob regardless of whether the value
  changed. Separately, a failed durable write during a knob set (disk full,
  encryption error) was silently swallowed and reported back as full
  success; the response now distinguishes a persisted set from one that
  only updated the live session, and a failed write raises a warning event.
• The per-turn memory refresh re-injected the entire
  session-start brief (identity, critical facts, topic communities, key
  memories) on every store advance instead of just what changed, so a busy
  session could re-pay thousands of tokens per turn for content the model
  already had. It now renders a bounded delta of only the records newer than
  the caller's own watermark, cross-session records carry an origin label, a
  burst larger than the render window still advances the watermark and marks
  the truncation, and a server-side debounce caps how often a session can
  fire the refresh without ever dropping a new record — it simply
  accumulates until the next allowed render.
• A retrieved memory was never marked as reviewed, so
  the nightly forgetting sweep and cluster-replay step both treated it as
  never used. Recall now stamps last_reviewed on every returned record as
  part of the write it already issues, independent of the reconsolidation
  dry-run flag and of that config loading successfully. Using a memory now
  protects it from the forgetting sweep's rolling window and lets nightly
  replay reach it.
• The monotropism-depth topic names did not survive a
  daemon restart: the map is persisted, but the in-process cache that the
  rank gain, the response reorder, and the duplicate check all read from
  stayed empty until a nightly cycle happened to run in that same process,
  so a restart silently disabled the whole feature until the next sleep.
  Boot hydration now loads the persisted topic names the same way it loads
  the rest of the behavioural profile, so the feature keeps working
  immediately after a restart.
• A topic's monotropism-depth learning could reset for
  no reason: when a topic's display name needed a disambiguating second
  word (two topics sharing the same top word), that second word was
  recomputed fresh every night, so it could rotate even though the topic
  itself never changed, starting that topic's accumulated depth over from
  zero under the new name. The disambiguating word is now sticky the same
  way the topic's main name already was, so a stable topic keeps its
  learned depth.
• Removed records could still be returned by recall. The two maintenance
  operations iai-mcp blob-quarantine --apply and iai-mcp idem-dedup --apply
  soft-deleted a record but did not mark it not-live in the same write, so recall
  (which serves only live rows) kept returning it, and the row also stayed
  resident in the warm caches. Both operations now clear the live flag together
  with the deletion timestamp and evict the removed ids from the exact-match and
  recency caches. A one-shot boot-time self-heal repairs a store already drifted
  by the old behavior, and a source-scan check keeps any future deletion path
  from reintroducing the split.
• Encryption key recovery resurrected deleted records. iai-mcp crypto recover-prior-key rebuilt every row through the fresh-insert serializer, which
  dropped the deletion timestamp, the live flag, and the pending-embedding flag —
  so recovery brought soft-deleted records back as findable and silently marked
  pending rows ready. Recovery now stages each row byte-for-byte from storage
  (preserving deletion and pending state) and re-keys only the encrypted columns
  under the new key. The recovery staging table is built from the canonical
  schema so the vector-index key and the record-id uniqueness constraint survive
  the swap. Stores already recovered under the old behavior cannot be repaired
  (the deletion timestamp was erased); this fix prevents it going forward.
• A duplicate identifier written straight to the memory
  table (bypassing the normal upsert path) silently appended a second row
  instead of being rejected, on the primary storage engine only — the
  fallback engine already refused it. Both engines now refuse a duplicate
  identifier the same way, and a routine batch write that hits a genuine
  duplicate no longer keeps retrying the same batch forever; the offending
  entry is dropped once its content is confirmed already stored, and
  anything else is surfaced loudly rather than assumed safe. Encryption key
  recovery also now refuses outright, rather than silently duplicating
  content, if it ever finds a source that already holds two entries under
  the same identifier.