Link 3.0.0

Added

• Recall works outside English. The tokenizer split on [^a-z0-9]+, so
  every non-Latin script produced zero tokens: Japanese, Chinese, Korean,
  Russian, Arabic, and Indic memories were unfindable, and nothing said so -
  recall simply returned nothing. Accented Latin fared little better,
  über Größe Straße became {ber, stra}, and déploiement could not be
  found by typing deploiement. Scripts written without spaces are now cut
  into character bigrams, Latin accents are folded (along with ß, ø, ı
  and friends), and the three-character floor - an English heuristic that
  erased shorter words wholesale - applies only to Latin text. Combining
  marks in Indic scripts are vowels rather than accents and are kept:
  stripping them turned मंगलवार into गलव. ASCII text keeps the original
  path exactly, so no existing memory changes token and no existing ranking
  moves; the LoCoMo track returns all nine published figures unchanged across
  1,536 third-party queries, and the ASCII path is marginally faster than
  before. Wiki page full-text search had the same bug one layer down: its
  query normalizer deleted every non-ASCII character before SQLite ever saw
  it, so lnk search for any non-Latin query produced zero terms. Both the
  index and the query side now segment text the same way recall does, and
  the FTS cache moves to v2 so existing indexes rebuild - the first recall
  after upgrading pays that rebuild once. Segmentation follows Lucene's
  CJKAnalyzer: overlapping character bigrams, no dictionary. The fast
  semantic tier's default model is English-trained; measured on a Japanese
  set it neither helped nor hurt (6/6 with and without), so non-English
  recall is lexical-quality by default. The multilingual static model
  (minishlab/potion-multilingual-128M, 101 languages) is 512 MB against
  the default's ~30 MB, so it stays opt-in:
  LINK_SEMANTIC_MODEL=minishlab/potion-multilingual-128M.
• lnk ingest for structured exports (#66, contributed by @jakobtfaber).
  A plan-first importer for supported structured sources, starting with the
  chezmoi-docs-graph-v1 adapter: provenance manifests hashed per output,
  staging through a temporary directory, validation before promotion, and
  explicit --replace-unmanaged and --prune gates so nothing is
  overwritten or deleted without being asked. Imported documentation lands in
  the wiki, never in memory, and is kept out of automatic personal-memory
  proposals. Review found two things, fixed before release: text-mode
  --apply crashed after succeeding, and the proposal guard matched a
  substring so a capture that merely mentioned the feature lost its
  proposals; it now keys off the export's shape.
• lnk stale - notice when a memory outlived the code it describes. The
  most repeated complaint about agent memory is that nothing can tell when a
  memory stopped being true: a note says a thing lives in a/b.py, the file
  is renamed, and the memory keeps being retrieved and believed. Hosted
  memory services cannot fix this because they never see the repository. Run
  lnk stale from inside a repository and it lists memories naming files git
  no longer has, with the successor path where git recorded a rename. A
  memory is questioned only when it names a path that is missing now and
  that git tracked before - without the second half, an unresolvable path is
  just prose and flagging it is the noise that teaches people to ignore the
  flag. The command changes nothing; findings go to the same review gate as
  everything else. Precision is measured rather than asserted:
  scripts/eval_staleness.py reports 0 false flags across 95 path references
  in this repository's own documentation, detects every probed deletion, and
  exits non-zero if either changes.

Measured and declined

• Usage-aware ranking. Four formulations were built and measured -
  additive frequency, tiebreak-only, recency decay in the Generative Agents
  form across the recommended 7-30 day half-life range, and an MMR diversity
  penalty - and none ship. Every one either made memories that had gone
  unread harder to find or did nothing; recency was worst, with the old half
  losing 0.0510 while the fresh half gained 0.0204 at a 30-day half-life. The
  reason is a category difference: those policies suit episodic observation
  streams, and Link stores durable constraints, which do not become less true
  for going unread - that is when they most need surfacing.
  scripts/eval_salience.py holds all four closed and fails on any
  regression, so the next attempt has to clear the same bar.

Fixed

• Memories with non-Latin titles no longer collide. Page names came from a
  [^a-z0-9] slug, so a Japanese, Hindi, Korean, Arabic or Cyrillic title
  slugged to nothing and every such memory was filed as memory.md. The
  second one then matched the duplicate gate's same-slug rule with a perfect
  score and was refused, so a non-English user could save exactly one memory.
  Titles now keep their own script (東京のデプロイ曜日.md, डिप्लॉय-का-दिन.md),
  Latin accents fold so filenames stay typeable (zurich-deploy-regel.md),
  the length cap counts bytes for multibyte scripts, and titles that slug to
  nothing (emoji-only) no longer read as duplicates of each other. ASCII
  titles keep their exact historical slugs; existing pages are untouched.
  The recall packet's ranking key uses the same rules, so non-Latin wiki
  pages no longer share one empty key.
• LinkBar 1.4.0. Fixes first: the health probes ran on the main
  actor, so the popover froze for a second or two on every Status refresh;
  they now run concurrently off it, and the five inbox reads run together
  instead of one after another. The review inbox showed five items and
  silently hid the rest; the tab now scrolls and stays on screen on a 13"
  display. Approve, archive, accept and discard confirm what they did, and a
  refused save reports the CLI's actual reason (duplicate, conflict, secret)
  instead of a guess. The live-agent pulse names the repository from the
  transcript's working directory, so link-pr66 no longer reads as pr66.
  lnk is found for pipx and venv installs, which a Finder-launched app's
  minimal PATH used to miss. Then the two things people asked for: the
  workspace is chosen in Settings and remembered (a Finder-launched app never
  saw LINK_WORKSPACE), and the memory filter matches the way Finder does,
  so typing zurich finds Zürich.
• LinkBar shows stale references. A Status row runs lnk stale against
  the repository your most recent agent session is working in and lists the
  memories that name files it no longer has, with a one-click filter on the
  Memory tab and an amber dot in the menu bar. On a CLI older than 3.0 the
  row says so instead of checking forever. The palette gained ↑↓ selection,
  and recall rows mark memories that default recall would hold back.

Removed

• The Bar CI investigation integration. Its Cloudflare side is gone, so
  the collector, summary poller, PR comment writer, their tests, and the
  workflow that drove them come out. The release-hygiene network allowlist is
  back to a single entry, the local viewer smoke test.

Changed

• The retrieval benchmark reports precision, not only recall. Recall is
  the number this category publishes, and it cannot separate a system that
  retrieves cleanly from one that returns everything, because returning
  everything scores 1.0 by construction. On the same 1,536 third-party LoCoMo
  queries, a whole-store dump carries 0.26% signal while Link's top-1 packet
  reaches 0.3086 precision on the fast tier - 117x - and pays a real recall
  cost that is published in the same table. The track now also reports the
  ceiling each cutoff allows (LoCoMo evidence sets average 1.53 turns, so
  precision@10 cannot exceed 0.152 for anyone) and R-precision as the
  k-independent figure to compare across systems.

Install / upgrade

brew upgrade lnk                              # or: pip install --upgrade link-mcp
brew upgrade --cask gowtham0992/link/linkbar  # macOS menu bar app, LinkBar 1.4.0

Thanks to @jakobtfaber for lnk ingest (#66).