Prepares for Reddit/HN/Discord push. No breaking changes.

Fixed

• SKILL.md (auto-invocation): 45% → 19% Japanese. English speakers now get auto-fires on natural EN phrases ('how did we solve this before?', 'same error again', 'remember this') in addition to JP triggers.
• install-skill CLI: shows bilingual test examples (was JP-only).
• session-extractor EN regex (for linksee-memory-import): adds let's go / pivot / switch to / settled on / approved / doesn't work / stuck / same error again / hit an error / debug / broke / revert. Also restores JP うまくいかない / ハマった.
• Caveat forget error hint: correctly distinguishes layer-protection (permanent) from pin-protection (releasable by lowering importance).

Unchanged

MCP protocol surface — all existing clients continue to work.

npm

npm install -g linksee-memory@0.2.0

Verified in English (this release)

• remember / recall / recall_file / update_memory / list_entities / read_smart / forget / consolidate — all work cleanly on EN content
• FTS5 matches EN tokens with relevance 1.0
• Layer aliases (warnings → caveat, decisions → learning, etc.) resolve in EN
• Caveat protection mechanism behaves as designed

Patch release based on dogfood feedback.

Change

Pin threshold lowered from importance >= 1.0 to importance >= 0.9.

Why

Memories set with importance: 0.95 or 0.9 were not being treated as pinned despite intent. The CHANGELOG said "importance=1.0 implicitly pins" but in practice agents and users naturally pass 0.9-0.99 to mean "this is important — please don't lose it". The strict 1.0 requirement created a UX surprise.

Effect

• All existing memories with importance >= 0.9 (including older entries set to 0.9 or 0.95) become pinned automatically.
• They are now exempt from the auto-forget sweep.
• recall and remember responses report pinned: true for them.
• No database migration needed.

Backward compat

100%. Anyone who was setting importance: 1.0 explicitly continues to work. The threshold is more lenient, not more strict.

Verify

npx -y linksee-memory@0.1.1
# or
npm install -g linksee-memory@0.1.1

One week of dogfooding surfaced several structural improvements. All changes are backward-compatible.

New tools

• update_memory — atomic edit preserving memory_id. Fixes the orphaned session_file_edits.memory_id bug when correcting a memory via forget+remember.
• list_entities — fast "what do I know about?" primitive. Returns layer breakdown per entity with kind/min_memories filters and pagination.
• linksee-memory-stats (CLI) — local DB summary. Shows entity count, layer distribution, top entities, top edited files. --json for scripts.

Recall improvements

• match_reasons: explicit array explaining WHY each memory was returned (content_match_fts / entity_name_match / heat:hot / pinned / ...).
• score_breakdown: per-dimension sub-scores.
• Pagination: offset / has_more / stopped_by.
• limit: hard cap complementing max_tokens.
• band filter: hot/warm/cold/frozen.
• mark_accessed=false: preview queries that don't bump heat.
• Layer aliases: decisions → learning, warnings → caveat, how → implementation, etc.
• FIX: opportunistic momentum refresh for entities recalled >1h after last remember().

Remember improvements

• Quality check: rejects pasted assistant output / CI logs / stack traces. Pass force=true to bypass.
• importance=1.0 = pin. Memory is exempt from forget auto-sweep even outside caveat layer.

Forget

• Pinned memories (importance>=1.0) preserved alongside caveat layer.
• Clear structured error for missing/protected memory_id.
• dry_run returns sample_ids_to_drop.

Consolidate

• dry_run=true preview — reports cluster count + candidates without writing.

Infra fixes

• migrate.ts: guard meta-table lookup for fresh DB.
• server banner now reports v0.1.0.

Install / upgrade

```bash
npm install -g linksee-memory@0.1.0

or

claude mcp add -s user linksee -- npx -y linksee-memory@latest
```

Links

• npm: https://www.npmjs.com/package/linksee-memory
• MCP Official Registry: io.github.michielinksee/linksee-memory
• Glama: https://glama.ai/mcp/servers/michielinksee/linksee-memory

Thanks to everyone who dogfooded v0.0.x and surfaced these rough edges — the module is meaningfully better as a result.

What's new

This is the first registry release of Linksee Memory. The package is now discoverable on multiple MCP registries and awesome lists.

Where to find linksee-memory

• npm: https://www.npmjs.com/package/linksee-memory
• MCP Official Registry: `io.github.michielinksee/linksee-memory` (API)
• Glama.ai: https://glama.ai/mcp/servers/michielinksee/linksee-memory
• LobeHub Marketplace: lobehub/lobehub#13912 (under review)
• punkpeye/awesome-mcp-servers: punkpeye/awesome-mcp-servers#4999 (under review)
• mcpservers.org: under review

Install

```bash
claude mcp add -s user linksee -- npx -y linksee-memory
```

Or with any MCP client that speaks stdio.

Features

• 6-layer structured agent memory (`goal` / `context` / `emotion` / `implementation` / `caveat` / `learning`)
• Token-saving file diff cache (`read_smart`) — measured 86% savings on typical TS edits, 99% on unchanged re-reads
• Cross-agent portability — single SQLite file at `~/.linksee-memory/memory.db`
• Active forgetting via Ebbinghaus curve; caveat-layer protected
• Optional Stop hook for Claude Code — auto-captures every session
• JP/EN trigram FTS5 search
• Optional anonymous opt-in telemetry (`LINKSEE_TELEMETRY=basic`)

Tools (6)

`remember`, `recall`, `recall_file`, `read_smart`, `forget`, `consolidate`

Verified

• Glama introspection check passed
• Dockerfile in repo root for reproducible container builds
• MIT licensed