v0.1.8

Event dating gains a provenance fallback, and the public contract stops
breaking older clients on additive result fields. CONTRACT_VERSION stays
0.1.0.

Added

• Tier 3 event facts may promote on mentioned_at -- the deterministic
  earliest date of a fact's source messages -- when no occurred_at is
  available, instead of sinking in Tier 2 forever. mentioned_at is derived
  provenance, never model output, and is never substituted into occurred_at
  (ADR 0037, Memory Invariant 11). It is exposed on LongTermFact and joins
  occurred_at ahead of created_at on every temporal filter and on the
  event-timeline sort.
• Light renders each message with a transient observed=<date> <Day> label so
  the extractor can resolve relative references. The label is prompt
  scaffolding and is never persisted (ADR 0038).
• Operator memory tooling in the CLI: vexic operator review-export and
  vexic operator rebuild-copy.
• Offline evaluation and diagnostic harnesses under scripts/: extraction
  prompt and Light time-context ablations, an oracle-evidence experiment, a
  class-3 gap probe and simulation, and a Deep-backlog replay. Plus
  preference rescoring (vexic.longmemeval_rescore) and rubric-aware judging
  for preference question types.

Changed

• Result models tolerate unknown fields. MemoryResult and the result-only
  payload models now derive from MemoryResultModel (extra="ignore"), so a
  client pinned to an older contract no longer raises ValidationError when a
  newer server returns an additive field -- which adding mentioned_at would
  otherwise have done to every pre-0.1.8 client. Requests and their payload
  models keep extra="forbid", because an unknown key in caller-written input
  is a typo and should still fail loud.
• Deep candidate subject keys normalize case and whitespace for dedup and
  histogram grouping (ADR 0039), and extraction prefers named entities as the
  subject.
• The rebuild-copy secret scan covers every column of every table rather than
  only text-declared ones, and refuses a source database missing a canonical
  table instead of silently creating an empty replacement.

Fixed

• An echoed observed= render label could persist into candidate text and
  have its recording date copied into occurred_at as event time. The label
  strip now keys on whether the date is one the window actually recorded, and
  a source-date backstop refuses any in-text date that is a cited message's
  recording date, whatever wording surrounds it.
• occurred_at no longer survives when it contradicts, or invents precision
  beyond, the single date the fact text states.
• mentioned_at is canonicalized on the promotion path and in Deep selection,
  so a value written by the canonical-migration importer or a host cannot
  become a durable temporal key that no as_of or event_before filter
  matches. Blank values are healable by the init backfill instead of sinking
  the candidate permanently.
• Candidates reduced to empty text by the marker strip are dropped and counted
  in the run's drop telemetry rather than staged with an empty FTS body.
• Diagnostic candidate reads no longer issue schema DDL, so a frozen run
  database opened read-only is readable, and a run database predating
  mentioned_at reports its real candidates instead of an empty set.
• The extraction ablation harness opens its frozen corpus read-only, so a
  WAL-mode evidence artifact is not rewritten on close.
• Preference rescore tolerates legacy run-database schemas, and the recorder
  hook fails fast rather than blocking when handed an interactive stdin.

v0.1.7

Recorder deadline, retry, and error-reporting reliability for the async Stop and SessionStart hooks. No public memory contract change; CONTRACT_VERSION stays 0.1.0.

Fixed

• Session-start prime reads fan out across parallel daemon workers under a single end-to-end deadline (default 20s), so a stalled hosted read is abandoned rather than waited into the SessionStart hook kill window and can no longer eat the whole injected block (#250).
• Recorder ingest bounds retries end to end: transport failures, 408, and 429 responses retry with a capped, jittered backoff and honor a bounded Retry-After, and the per-attempt socket timeout is capped to the remaining budget. The default socket timeout composes with the ingest deadline to stay inside the Stop hook kill even when an un-preempted in-flight read overshoots the deadline (#252, #253).
• MCP tools/call error path sanitizes pydantic ValidationError messages through the client-safe builder instead of echoing raw input values (#251).

Changed

• fetch_prime_context returns a PrimeFetchResult (context plus per-leg timing/outcome) instead of a bare str; recorder prime writes phase: "started"/"finished" records into a sibling -prime.json status file so overlapping Stop-hook ingests and prime cannot erase each other's records (#250).
• Product glossary adds the dreaming and recorder terms; the hosted-mvp sweeper section notes the bounded prelude retry.

Full Changelog: v0.1.6...v0.1.7

v0.1.6

Session-start priming robustness plus recorder and hosted-layer reliability
fixes. No public memory contract change; CONTRACT_VERSION stays 0.1.0.

Fixed

• Session-start prime usage guidance survives truncation: a fixed framing
  block leads the injected snapshot and footer space is reserved before
  capping, so the instructions can no longer be the first thing truncated
  away (#246).
• Recorder transcript-cursor writes are monotonic, so overlapping async Stop
  ingests can no longer regress the cursor and re-post already-ledgered
  rows; equal-offset content corrections still write through so same-length
  transcript rewrites self-heal (#248).
• Hosted write-route preflight 400 responses sanitize pydantic
  ValidationError messages through the same client-safe builder as the
  HTTP adapter, instead of echoing raw input values (#249).
• Dream-phase prelude retries retryable operational faults with a bounded
  backoff (#243).
• Recorder Stop hook fails open on transient hosted errors instead of
  derailing the conversation with a blocking exit (#244).
• Hosted 400 responses in the HTTP adapter sanitize pydantic
  ValidationError messages (#245).

Changed

• Prime snapshot sections carry per-item caps (transcript hits and the
  recap body) with an explicit truncation marker at both fetch and render
  time, so one long item cannot starve the other sections; the full text
  stays reachable through the recall tools (#246).
• MCP server_instructions now describe session-start priming: visible
  snapshot facts need no re-search (explicitly subordinate to the
  search-before-denying rule) and truncated snapshot items are flagged as
  incomplete (#246).

Full Changelog: v0.1.5...v0.1.6

v0.1.5

Reliability patch plus evaluation tooling. No public memory contract change;
CONTRACT_VERSION stays 0.1.0.

Added

• LongMemEval weighted sampling (--type-weight TYPE=N) and the
  vexic.longmemeval_analysis miss-classification module for diagnosing
  retrieval misses in eval runs (#237).

Fixed

• Transcript ingest strips harness-injected <task-notification> blocks the
  same way as system-reminder blocks, keeping subagent reports and tool
  returns out of searchable transcript text (ADR 0034, #235).
• Transcript recall uses any-token OR FTS semantics with bm25 ranking, so a
  query only partially matching stored text still recalls it (ADR 0036, #236).
• Writes to a tombstoned scope fail closed both before and after physical
  purge (ADR 0022, #240).
• Four hosted-layer alpha staging shortcuts are hardened in the in-process
  hosted shell (#239).

Changed

• GitHub Actions workflows moved off deprecated Node 20 runtimes, are
  SHA-pinned, and Dependabot now watches the actions ecosystem (#238, #241).

Full Changelog: v0.1.4...v0.1.5

v0.1.4

Reliability patch for the internal-alpha hosted stack. No public memory
contract change; CONTRACT_VERSION stays 0.1.0.

Fixed

• SessionStart priming no longer discards all memory context when a hosted
  read times out mid-response: read-phase transport and decode failures are
  normalized into the degradation path, so prime emits a partial context
  instead of nothing (#231).
• Canonical migration artifacts tolerate additive schema columns and the
  import row loop is one atomic transaction whose rollback failures preserve
  the original import error (#230); a raw libsql DSN string import target
  fails closed on host-owned extension tables instead of skipping the
  pre-import guard (#233).
• Control-plane schema migrations are concurrency-safe (#226), and
  control-plane retirement cuts live access (#227).
• Dream-phase failures are durably recorded and gate the retry clock (#222),
  dream sweep state is stamped with job-completion time (#225), withheld-stamp
  dream chains back off briefly after failures (#224), and the summarize
  watermark holds when SUMMARIZE never ran (#223).
• Light extraction drop counts are durable telemetry, and all-dropped runs
  report partial instead of success (#229).
• Usage is captured under the pydantic-ai property form, and missing usage
  fails loud instead of silently reporting zero (#228).

Full Changelog: v0.1.3...v0.1.4

v0.1.3

Reliability and hardening patch for the internal-alpha hosted stack. No public
memory contract change; CONTRACT_VERSION stays 0.1.0.

Added

• Remote libSQL queries are bounded by a wall-clock deadline (#213).
• Recorder keeps a local transcript cursor, so the Stop hook no longer rereads
  the full transcript each turn (#208).
• Per-tenant Turso token cache is bounded with LRU eviction (#207).

Changed

• Transient Turso connect faults are absorbed at the connection boundary (#219).

Fixed

• Light extraction drops miscited candidates instead of failing the whole
  batch (#201), handles assistant-heavy task transcripts (#211), and filters
  Claude Code harness envelopes out of ingested transcript (#212).
• The dream in-flight lock is a durable control-plane lease; a cancelled dream
  job holds its lease instead of freeing a live scope (#203).
• Every reasoning agent gets output headroom rather than output-sized caps
  (#199).
• Operator CLI honors VEXIC_CONTROL_PLANE_TARGET, so revoke-key revokes on
  production (#206).
• Release paths from PR #214 hardened (#215).

Full Changelog: v0.1.2...v0.1.3

v0.1.2

Feature and hardening patch for the internal-alpha hosted stack. The public
memory contract gains one additive operation; CONTRACT_VERSION stays 0.1.0.

Added

• load_active_context, a structured sibling of fresh_context that returns
  serialized transcript messages a host can replay as model message history
  instead of rendered priming text.
• LongMemEval evaluation harness rehomed into vexic.longmemeval with a CLI,
  host-port judge wiring, and an --allow-live provider-call gate (COA-342).
• Hosted control-plane catalog can route to Turso behind a flag, with a
  migrate_control_plane migration path (COA-360).

Changed

• Dream sweeper storage routes through the customer-target resolver, and raised
  dream-phase output token caps with surfaced sweeper logs (COA-352, COA-355).
• Hosted ingest storage ValueErrors are classified as 5xx and surface only a
  stable error code (COA-356).

Fixed

• Dream sweeper state writes lost to reaped Turso Hrana streams are now retried
  (#194).

v0.1.1

What's Changed

• chore(ci): add gitleaks full-history secret scan by @PyroDonkey in #158
• Release: gitleaks secret-scan CI gate by @PyroDonkey in #159
• docs(agents): document live_retrieval_baseline harness owner by @PyroDonkey in #161
• fix(hosted): harden control-plane against mass deletes (COA-320) by @PyroDonkey in #160
• docs: harden repo docs for public release by @PyroDonkey in #162
• Release: docs public-release hardening + COA-320 control-plane by @PyroDonkey in #163
• docs: promote AGENTS.md to repo root as single source of truth by @PyroDonkey in #164
• chore(release): bump version to 0.1.1 by @PyroDonkey in #165
• release: 0.1.1 by @PyroDonkey in #166

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Initial public release of the Vexic memory core.

Added

• Local-first, provenance-first memory core for long-running AI agents.
• Public contract models (MemoryScope, MemoryCapability, MemoryService)
  with a versioned contract surface.
• LocalMemoryService, a SQLite reference implementation with sqlite-vec
  vector search and hybrid retrieval.
• Three-tier memory model: append-only cleaned transcript, staged memory
  candidates, and durable long-term facts, each carrying source provenance.
• Read-only local MCP server exposing recall_conversation_history and
  recall_user_memory.
• Conformance and reliability test suite.