Skip to content

Link v1.6.0

Choose a tag to compare

View all tags
@gowtham0992 gowtham0992 released this 10 Jul 00:14
v1.6.0
e99e933
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Link 1.6.0 makes the memory loop automatic: your agent no longer has to remember to use its memory. Everything stays local, plain Markdown, and review-gated — nothing is saved durably without your approval.

Automatic memory (session hooks)

lnk connect <agent> --hooks installs session hooks for Claude Code, Codex, and Cursor:

  • Every new session starts with a bounded Link memory brief injected automatically.
  • Session end stores proposal-only notes with memory candidates for your review.
  • Noise controls: sessions with nothing memory-worthy are skipped, duplicate end events are deduplicated, and proposals are mined from your words only — the assistant's prose is never attributed to you as a preference.

Hook writing is idempotent and non-destructive: existing hooks and settings are preserved, and reruns replace only Link's own entries.

Hybrid semantic recall (opt-in)

pip install "link-mcp[semantic]" + lnk semantic --setup adds a small local embedding model so recall finds memories phrased in completely different words. Lexical recall stays the default and the fallback.

  • Local-first guarantee holds: the model loads offline-only (only the explicit --setup may download it once), embeddings are plain JSON under .link-cache/, no vector database, and LINK_SEMANTIC=off disables the layer.
  • Two tiers: a fast static-model tier for instant CLI/hook use, and link-mcp[semantic-quality] (all-MiniLM-L6-v2, local ONNX) preferred automatically when installed.
  • Honest labels: every recalled memory carries match: lexical | semantic | hybrid, and a match with no lexical evidence is capped at moderate confidence.

Measured, with regression gates in CI

  • On the bundled 1,176-case benchmark: hybrid recall lifts token-overlap hit@1 from 0.589 to 0.703 and doubles-to-triples zero-token-overlap paraphrase hit@3/hit@5, at ~2.8 ms per recall in-process. The quality tier roughly quadruples pure-paraphrase hit@3/hit@5 over lexical.
  • On the third-party LoCoMo long-term-memory dataset (1,536 queries, retrieval stage only, no LLM anywhere): any-evidence hit@10 0.578 → 0.685.
  • Methodology, hardware, limitations, and failed ablations: benchmarks/RESULTS.md.

Also in 1.6.0

  • lnk consolidate: a read-only backlog plan (pending captures, duplicates, review queue) with paste-safe commands — nothing is merged or deleted without approval.
  • First-ten-minutes fixes from cold-walking Link as a new user: lnk onboard now leads with the automatic-memory path, an empty recall explains that paraphrase matching is opt-in, and python -m link_mcp --help prints usage instead of silently starting the server.
  • The README and Getting Started page show the loop end-to-end, recorded from real commands.

Install / upgrade

brew install gowtham0992/link/link     # or: brew upgrade link
pip install --upgrade link-mcp         # MCP server
lnk onboard                            # guided setup, prints the --hooks command

Full changelog: CHANGELOG.md

Assets 2
Loading