SkillFoxx

Link MCP

Local memory for AI agents in plain Markdown files: it keeps only what you approve and serves it through MCP

  • 218 ★ on the repository
  • Last commit
MCP server

Extended access

Why this level

  • The server writes Markdown memory files and reads project files to check memories about code
  • Rules inside memories can block tool calls, so approve them carefully
All reasons and checks

gowtham0992/linkAuthor:gowtham0992

Install

Manual install

brew install gowtham0992/link/link
lnk setup

Homebrew install and wiring of every detected agent; without Homebrew use `pip install link-mcp`.

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • The server writes Markdown memory files and reads project files to check memories about code
  • Rules inside memories can block tool calls, so approve them carefully

This is third-party code. Review the repository files before installing.

What it does

Link is a memory layer for local AI agents. Notes, transcripts and documents from the `raw/` folder become an inspectable Markdown wiki in `wiki/`, and explicit remember-this requests become memories you review and approve. Agents receive compact context packets with provenance through the CLI, MCP, skills or a local viewer. Codex, Claude, Cursor, Kiro, VS Code, Copilot and Antigravity read the same memory instead of keeping a separate profile at each vendor. An approved memory can carry a rule such as `ask command: git push --force` or `deny write: migrations/**`, and it is checked before a tool call with no model involved. Memories about code are re-read against the repository and flagged when the code moved on.

Who it is for. People who work across several agents and are tired of re-explaining preferences, repository decisions and constraints to each one.

Good fit when

  • You say deploy on Tuesdays once, approve the entry and another agent remembers it in a new session
  • An agent must never run `git push --force`: a rule in an approved memory stops the call or asks for confirmation
  • Old memories about files and scripts are checked for freshness after a large code rework
  • Memory accumulated in Claude Code has to move to a shared store with `lnk import claude-code`

Not a fit when

  • You need a semantic knowledge graph with model-based entity extraction: Link has no model in the write layer
  • Memory has to sync across machines over a network: by the project description it is stored locally
  • You want fully automatic memorizing with no review: writes to memory go through your approval

Example request

Remember that we deploy on Tuesdays and never force-push to the main branch, and show me the entry for approval

MCP

Transport
stdio
Authentication
not required

Limitations

Memory is stored locally in Markdown and the project's CI blocks outbound network code in the runtime. Optional semantic models load once through an explicit setup and then work offline. Writes to memory go only through your approval, so the agent proposes and you confirm. On macOS there is a menu bar app, `linkbar`, installed with Homebrew. The LoCoMo and LongMemEval benchmarks in the README were measured by the author on their own setup. The project is MIT-licensed and the package is called `link-mcp` on PyPI.

How to disable. Remove the server registration from your clients (`claude mcp remove link`) and the memory folder if it is no longer needed. The files are plain Markdown and are deleted by hand.

FAQ

How do I wire Link into all my agents?

Install the package and run `lnk setup`: it wires every agent it finds. You can check the result with `lnk status`.

Where is the memory stored?

On your computer in plain Markdown files you can open, grep or git-diff. Nothing goes to the cloud.

Can I import existing memory?

Yes, `lnk import claude-code` moves the memory accumulated in Claude Code into the shared store.

What does the agent see at session start?

Compact context packets with provenance. `lnk receipt` shows which memories reached the session and whether they arrived whole.

How is Link different from Mem0 and Letta?

There is no model in the write and recall layer: ingestion and recall are deterministic and memories are readable Markdown files instead of vectors or a graph.

Discussion

No questions or reviews yet.

MCP servers

Long-term memory for AI agents: Markdown files as the source of truth, a SQLite index as a cache, shared by Claude Code, Codex and others

CLIExtended access4.3K

ai-memory retains session knowledge in a Markdown wiki, exposes MCP retrieval and passes context between agents and machines.

MCP serverExtended access15.7K

Long-term agent memory in local Markdown files with a link graph, semantic search and MCP access

MCP serverExtended access4.6K

The awiki CLI keeps agent memory in a markdown vault: search before the web and save whatever is worth keeping.

CLIExtended access56

Local memory for coding agents: sessions turn into a git-versioned Markdown wiki and come back through an MCP server.

MCP serverExtended access130

Git-native project memory for AI agents: Markdown inside the repo, local recall with no model call, and gated delivery.

CLIExtended access426

README in short

The text opens with a promise of shared local memory for different agents, shows a quick install, explains five working steps and a problem and answer table. Next come five architectural commitments, benchmarks, the tool list and setup.

README badge

Report a problem