Link

Link gives AI agents persistent memory so you don't have to repeat preferences, project decisions and context in every session. Multiple agents can read the same local Markdown files, which you can inspect in an editor or Obsidian.

• Reviewed memory: Agents propose memories for your approval, including captures from session-end hooks; session-start hooks supply saved context to the next conversation.
• Source-linked wiki: Agents turn notes, transcripts, articles, PDFs and screenshots into linked pages whose claims point back to the original material.
• Compact recall: Retrieve relevant context with source references and confidence labels; optional local semantic matching finds memories even when a query uses different words.
• Memory history: Review dates bring stale facts back for checking, expiry removes them from default recall, and dated queries retrieve what was recorded as true at an earlier point.
• Controlled sharing: Sync reviewed memory across machines or teams through a Git remote you control, with secret scanning, private captures excluded and conflicts sent for review.

Link is open source under MIT and runs on Python. It integrates with Claude Code, Codex, Cursor, GitHub Copilot and other agents through MCP, skills and editor connections. The default memory layer needs no LLM, API or vector database. Recall works offline; optional semantic models require an initial setup before offline use. Memory stays local unless you choose to sync it.

A local web viewer provides workspace browsing, and LinkBar adds recall and review from the macOS menu bar. Local retrieval records show which memories agents read and when, without storing the questions or syncing the records.