OwnMem 0.6.0

[0.6.0] - 2026-09-09

Breaking

• Events record surface as hook rather than claude-hook, and carry two new
  top-level fields: agent (which coding agent produced the row) and nested
  (more than one host marker was present, so attribution is refused rather than
  guessed). Rows written before this release keep claude-hook and are reported
  as a separate legacy bucket; nothing is re-attributed after the fact, because
  those days genuinely mixed hosts and a guessed attribution is worse than none.
  Anything reading the event files or the report's buckets has to be updated.
• ownmem init --update installs the unattended daily pass on SessionStart
  and Stop, and generates <memory-dir>/git-hooks/post-commit as a third
  trigger. That pass commits the telemetry packages it writes and points
  core.hooksPath at that directory while the setting is empty and the hook is
  present. Both are new behaviour for a 0.5.x installation and both are
  disclosed at install time; set telemetry.auto_commit or
  telemetry.manage_hooks_path to false in <memory-dir>/config.json, or
  pass ownmem daily --no-commit, to switch them off. Not upgrading changes
  nothing: 0.5.x hooks stay as they are.
• The hook configuration is now v2 and the installed version is recorded in
  <memory-dir>/config.json. Run ownmem init --update once. v1 entries are
  replaced where they stand rather than appended beside, and hooks written by
  hand are left untouched. init --check and the daily pass both report an
  installation that is still on v1.

Added

• ownmem archive [--day] [--backfill] reduces a finished UTC day of local
  telemetry to a counted package under <memory-dir>/telemetry/, small enough
  to commit. Counts only: no query text or digest, topic body, path, machine or
  account name, or timestamp finer than the day. Idempotent, and a day with
  nothing new is not rewritten.
• ownmem report --since 7d --fleet merges the committed packages from every
  machine, leads with which machine covered which day, and names the days that
  have commits but no package. Overlapping days are counted once.
• ownmem daily is the unattended pass behind both of those: archive, commit,
  inspect what is already in Git, and print only what is wrong. It runs real
  work at most once per UTC day and always exits 0, so it can sit on a session
  hook and a post-commit hook without ever blocking either.
• Codex hook support: ownmem init --hook writes <project>/.codex/hooks.json
  with the same events as the Claude configuration. Reaching those hooks takes
  three separate permissions -- hooks = true under [features] in
  ~/.codex/config.toml, the project trusted in that same file, and the hook
  trust prompt on first sight -- and an untrusted project is silent, because it
  never discovers the file at all. init names all three, and reports a
  checkout that is missing from the project trust list. A Codex hook process
  receives no environment of its own, so every command declares its host
  explicitly and finds the checkout from its working directory.
• Grok CLI is a supported host (--hosts grok). It reads the Claude files
  through a compatibility layer, so it shares that configuration; init --check
  reports when the checkout is missing from grok's trusted-folder list.
• Agent attribution: recall, command, and turn events say which host produced
  them, and the report buckets by agent instead of merging three agents into one
  average.

Fixed

• The daily pass points core.hooksPath at <memory-dir>/git-hooks/, which
  ownmem init generates, and leaves the setting alone until that directory
  holds a post-commit hook. A path in this project's own checkout was used
  during development; in any other repository it named a directory that does not
  exist, and a core.hooksPath pointing at nothing installs no hooks at all --
  every hook that repository had would have stopped running, silently.
• Repository root resolution prefers the git checkout root when it carries an
  installation, so a nested memory directory copy -- a fixture, an exported tree
  -- inside a subdirectory no longer claims the root and receives that
  checkout's local data. A checkout without an installation still yields to the
  nearest one below it, so a monorepo package that installed its own memory
  keeps answering for the directories under it.
• PostToolUse and Stop hooks now run. v1 registered ownmem posttool and
  ownmem stop, which are subcommands of ownmem hook and not of the CLI, so
  both answered "unknown memory command" and every installation collected
  command outcomes and turn corrections from nowhere. v2 registers
  ownmem hook posttool and ownmem hook stop.
• The hook daemon keys its resident runtime per agent, so two hosts working in
  one checkout no longer share one connection and one attribution.
• ownmem init on the default core layer no longer accepts --hook and then
  installs nothing: asking for hooks now installs the layers they need.
• init --check no longer reports missing forever on a repository that
  replaced the synthetic example topic it was told to replace, and --update no
  longer puts that topic back into a governed corpus.
• The host set chosen at install time is remembered instead of re-detected.
  Grok leaves no marker in a repository, so an installation that named it
  reported drift on the next check.

OwnMem 0.5.5

Fixed

• A filtered test rerun no longer shrinks a file's roll-up in the test-execution ledger. Selecting one case of a forty-case file derived a roll-up of one, and because the latest record per key won, a memory citing the suite then read passed on the strength of that single case. On an unchanged file a passing roll-up that observed fewer cases now leaves the wider record standing; a failure, a wider run, or a change to the tested file still lands.
• ownmem health no longer prints 0 query(s), 0.0% abstained for an install that never wrote a usage log. The usage log is opt-in (recall --usage-file), so its absence is reported as unavailable rather than as a measurement.

OwnMem 0.5.4

Added

• swift-test is a known test runner in the test-execution ledger, for
  projects whose tests run through SwiftPM's swift test --xunit-output.
• ingestTestRun accepts any runner the caller supplies an adapter for, so a
  project can connect a toolchain of its own without editing the known list.

OwnMem 0.5.3

Fixed

• Trust receipts no longer bind a test anchor to a file inside a nested checkout
  (a linked git worktree such as .claude/worktrees/<agent>/, a submodule, a
  nested clone). The lookup by suite name descended into those copies, .claude
  sorted ahead of the real directory, and once the worktree was cleaned up the
  memory was quarantined as missing-path with its test untouched in the main
  tree. Any directory carrying its own .git entry is now skipped, and the path
  a memory declares next to its test locator takes precedence over the
  repository-wide lookup.
• Directory-anchor fingerprints skip nested checkouts for the same reason.
• A suite-level test anchor, whose symbol is the file's own type name, now
  resolves to that file's roll-up in the test-execution ledger instead of
  reading never-run forever: no runner reports a case by the suite's name.

OwnMem 0.5.2

Patch release so host plugins stop teaching a frozen protocol. Skills now route to npx ownmem instead of hard-coding recall-feedback flags, and the dashboard skill no longer writes a long-lived access token under /tmp. Negated high-risk recall reads the receipt authority ceiling so evidence drift cannot leak a normative R5 prohibition. Plugin docs now state that third-party Claude Code auto-update is off by default and that npm install ownmem@latest plus npx ownmem init --update remains the version path.

OwnMem 0.5.1

Patch release for truthful recall-feedback diagnostics. Reports now distinguish wrong deliveries from wrong abstentions, preserve legacy unpaired verdicts as unknown, and publish the closed ownmem-report/v8 contract. Release verification passed repository tests, the 100-iteration 40-language benchmark, and npm pack dry-run.

OwnMem 0.5.0

Highlights

• Faster exact identifier, path, and error recall by narrowing candidate scoring before ranking.
• Lower allocation pressure during full index compilation; audited 1,000-topic P95 improved from about 2.4 seconds to about 1.9 seconds.
• Query-span and no-answer observability for safer evaluation without exposing query text.
• Versioned runtime and query semantics so benchmark and report windows remain attributable after upgrades.
• Refreshed README content in English, German, Spanish, French, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, and Traditional Chinese.
• Clearer product wording and discovery metadata for AI coding agent memory, project memory, and supported hosts.

The release benchmark covers 40 languages and 25 scripts with 128 positive and 40 negative queries. Recall@1, Recall@5, MRR, and abstention precision passed at 100% in the release gate, with zero model or network calls.

Upgrade

npm install -g ownmem@0.5.0
npx ownmem init --update
npx ownmem audit
npx ownmem compile

This release uses query schema v5 and runtime semantics v0.10. No manual data migration is required.

See CHANGELOG.md for the complete change list.

OwnMem 0.4.0

Added

• correct_abstain records the other way recall can be right: it returned
  nothing, and nothing was the right answer. correct rejects an empty result
  set, so until now a correct abstention could only be filed as wrong or
  coverage_gap -- recall's successes were being counted as its failures, and
  every operator's own error rate was inflated by exactly that much. It judges
  recall's behaviour; coverage_gap judges the corpus, and both can be true of
  one query: a question about an unrelated product is correct_abstain and no
  gap at all, while a real question this repository has never written down is
  both. It never takes --expected, and it is refused unless recall did in
  fact return nothing.
• The feedback reader exports FEEDBACK_CORRECT_VERDICTS. Three call sites
  independently spelled "still needs a person" as verdict !== 'correct', so a
  second correct verdict would silently have been counted as outstanding work
  in all three.
• readMissDismissalReceipts reads an ownmem-miss-dismissal-receipt/v1
  ledger, and pendingTriggerBackfills now excludes a miss recorded in it
  rather than re-proposing it every run. A dismissal is deliberately not a
  resolution: a resolution asserts the miss stopped reproducing, while a
  dismissal only records that a person looked and decided no trigger edit can
  reach it, so a queue with no terminal state re-grades that decision forever.
  A receipt without a reason is rejected. The reader and the exclusion ship
  first; no CLI command writes this ledger yet.

Fixed

• A rejected telemetry event no longer leaves its feedback row behind. The
  event is validated against a closed schema and can throw, and it used to be
  built after the row had already been appended: the operator reads an error,
  re-runs the command, and the ledger now holds the same verdict twice.
• lib/features/trust.mjs refuses to run as a direct entry point instead of
  loading, exporting everything and exiting 0 having done nothing at all. A
  clean exit from a module that needs the defaults its wrapper supplies reads
  as a successful run: two such invocations reported success while the ledger
  stayed stale.

Migration

• ownmem-recall-feedback/v3 gains one enum value rather than a new schema id.
  A 0.4.0 reader accepts every 0.3.0 ledger unchanged, but a 0.3.0 reader
  rejects a row carrying correct_abstain as invalid. Upgrade every reader in
  a repository before writing the new verdict into its ledger.

OwnMem 0.3.0

Full Changelog: v0.2.0...v0.3.0

OwnMem 0.2.0

Added

• A structural regression gate checks the root allowlist, local documentation
  links, ESM imports, canonical schema URLs, npm publication boundary, and
  cross-host skill mirrors.
• ownmem embed exposes the optional embedding workflow through the stable CLI
  instead of requiring a package-internal script path.

Changed

• Runtime entry points now have explicit boundaries: bin/ownmem.mjs is the
  executable, lib/index.mjs is the package API, feature orchestration lives in
  lib/features/, and runtime contracts live under domain-grouped schemas/.
• Tests and locked benchmark evidence moved to test/ and benchmarks/; npm
  now publishes an explicit runtime allowlist instead of every root-level MJS
  and JSON file.
• Community health files moved to .github/; translated READMEs, project docs,
  and media moved to docs/, reducing the GitHub root from 67 entries to 22.
• CI now uses five targeted runtime/platform combinations instead of repeating
  the full release gate across a nine-job Cartesian matrix. Superseded runs are
  cancelled, and official Actions are pinned to reviewed release commits.

Fixed

• The structure gate now reads the executable bit for bin/ownmem.mjs from
  Git's cross-platform index metadata. Windows no longer fails because its
  filesystem does not expose POSIX execute bits through fs.stat().

Migration

• Invoke runtime behavior through ownmem <command> and import the supported
  API from ownmem; root-level memory-*.mjs package internals no longer ship.
• Resolve JSON contracts through ownmem/schemas/*. The three schemas with
  canonical GitHub URLs now identify their real paths under schemas/.