⚡
read_smart
Token savings, measured.
AST- and heading-aware diff cache on file reads. 86% saved on a typical TS edit, 99% on unchanged re-reads.
◆Handover + drift detection · MIT
Hand your project over — with the reasons attached.
To your next session, to Cursor or Codex, to the person after you. Record a decision once and it comes back before the agent acts on it. Underneath: drift detection across every project you run — which one quietly left its goal, which was checked and holds, which nobody has looked at. Local-first, for Claude Code, Cursor, Codex and Gemini.
No account. No cloud. One SQLite file on your machine.
up to 99%
Tokens saved on re-reads
6
Memory layers
0
Cloud dependencies
Watch the map catch drift
linksee-memory map · live
Three pillars
Built on three things nobody else combines.
🔗
cross-agent
One SQLite file. Every agent.
Claude Code, Cursor, Windsurf, OpenAI Codex, Gemini CLI — all share one brain at ~/.linksee-memory/memory.db. No cloud, no account, no lock-in.
🧠
WHY-first
Six layers, not a flat log.
goal / context / emotion / implementation / caveat / learning. Forgetting is tuned per layer — caveats are protected.
7 tools + a product map
Memory is the entry point. The product map is the new surface.
Seven tools, two pillars and one on request. Memory: remember (content only — add anchor: {} to enforce a decision), recall (no arguments = the session brief; where = your position on the map and its blast radius), read_smart. Drift: drift_status, declare_anchor, resolve_drift. On request: share_draft (makes a share link only when the user asks). Nothing else is exposed — fewer tools behave the same across Claude, GPT, Cursor, Codex and Gemini. Plus the full MCP protocol surface: Tools + Resources + Prompts + Sampling + Roots + Elicitation.
🗺️
Map
recall({ where }) + linksee-memory map.
A map.yaml (git source of truth) describes how value reaches your user; the CLI answers where am I, what did this file promise, what does the code do, and what else must change — with file:line evidence. Add --lang ja for Japanese.
📚
Resources (NEW)
Memory as @-mentionable URIs.
memory://stats / memory://hot / memory://recent / memory://caveats + 3 templates (memory://entity/{name}, memory://layer/{layer}, memory://memory/{id}). Browse memory in your IDE without a tool call.
📝
Prompts (NEW)
5 reusable templates.
summarize-session / extract-caveats / weekly-consolidation / recall-and-write / entity-handoff. Memory-before-action discipline, baked in.
🤖
Sampling (NEW)
Server asks your LLM.
Auto-consolidation on startup can optionally ask the client's LLM to rewrite cluster summaries. Falls back to the heuristic when the client declines. No server-side LLM key needed.
📂
Roots (NEW)
Auto-scope to your workspace.
recall{path, scope_to_roots:true} reads the client's working roots and filters file history matches. Skips silently if the client doesn't expose roots.
❓
Elicitation (NEW)
User-in-the-loop on destructive ops.
remember{forget:true, memory_id:N} asks the user to confirm via the client UI before deleting. Newer MCP primitive — graceful degrade on clients without support.
The WHY structure
Six layers that make retrieval explainable.
Every memory is tagged with exactly one layer. Recall composites relevance × heat × momentum × importance and returns match_reasons so you — and the agent — know why a row came back.
🎯
goal
goal
The outcome the user is working toward. Persists across sessions so the agent doesn't drift.
🧭
context
context
Why this, why now. Surrounding constraints and stakeholders that shape the right answer.
💗
emotion
emotion
Tone cues from the user. Frustration, excitement, exhaustion — signals that change how to reply.
🛠
implementation
implementation
How something was done — including what failed. The difference between a clean diff and a scar.
⚠️
caveat
caveat
The "never do this again" drawer. Protected from auto-forgetting, always surfaced on recall.
🌱
learning
learning
Patterns distilled over time. Auto-consolidation compresses cold memories into protected learning entries on startup.
Drift detection
Forgetting is bad. Drifting is worse.
Last week you decided "FTS5, not vector search." This week a fresh session installs pgvector — nobody's fault, the context just died. Memory tools record what you did. Nothing watches when reality diverges from what you decided. That's the layer Linksee adds.
A drift is a question, not a verdict — "which one is correct now?"
📌1 · declare
Anchor your decisions.
"We use FTS5" / "never ship without tests" — declared explicitly, never mined from chat. 4 species: hypothesis, constraint, commitment, source-of-truth.
📡2 · watch
Reality is reconciled against intent.
Committed code is checked against your anchors. Four states: 🔴 drift · 🟡 review · ⚪ held · 🔵 aligned. Recorded evolution stays quiet — only unaccounted gaps fire.
🛡️3 · guard
Re-injection, before the action.
A hook re-surfaces the locked decision before the agent runs the violating command. It asks first; it only blocks the rules you explicitly hardened.
Resolve in conversation: fix the code / the direction changed / false positive — plus two gates: harden (block) and soften (warn).
Terminal — any MCP client
drift_status · local
> what's drifting?
28 anchors: 🔴 1 drift · ⚪ 1 held · 🔵 26 aligned
🔴 DRIFT"We use FTS5, not vector search"
reality: pgvector added in db/schema.ts — no recorded resolution
→ which one is correct now? [fix] [direction changed] [false positive]
🛡 guard: held `npm i pgvector` — anchor #7 re-injected before action
Three spells to remember
"use linksee"recall before acting
"linksee this"save it now
"what's drifting?"reconcile reality vs decisions
Get started in 5 minutes
Not a developer? Start here.
We'll walk through everything — installing Node.js, setting up Claude Desktop, and verifying it works. No terminal experience required for Claude Desktop.
Claude Desktop (recommended for non-developers)
Claude Desktop uses a settings file. We'll add linksee-memory to it with one paste.
- 1
Open Claude Desktop settings
Launch Claude Desktop → Settings → Developer → "Edit Config". A JSON file opens in your editor.
- 2
Paste the linksee-memory block
Inside "mcpServers": { ... }, add the linksee key shown below. If mcpServers already has entries, add a comma before the linksee line.
"linksee": { "command": "npx", "args": ["-y", "linksee-memory"] } - 3
Restart Claude Desktop
Fully quit and reopen (not just close the window). linksee-memory loads on the next launch.
- 4
Verify it works
Start a new chat and type: "Can you remember that I prefer TypeScript over JavaScript?" Claude should confirm it saved this to linksee-memory.
How it compares
The features most agent-memory tools quietly skip.
| Feature | Mem0 / Letta / Zep | Claude auto-memory | linksee-memory |
|---|---|---|---|
| Cross-agent (Claude + GPT + Gemini) | △ | ✕ | ✓ |
| WHY-layered structure | ✕ flat | ✕ flat | ✓ 6 layers |
| AST-aware file diff cache | ✕ | ✕ | ✓ |
| Active forgetting (curve-based) | △ | ✕ | ✓ |
| Local-first · no account required | ✕ | ✓ | ✓ |
| match_reasons on recall | ✕ | ✕ | ✓ |
Give your agent a brain it can trust next week.
MIT-licensed. One binary. Stays on your machine.