Favicon of Link

Link

Local AI agent memory stores reviewed context as Markdown, works with Claude Code, Codex and Cursor, and recalls it offline under an MIT license.

Screenshot of Link website

Link is an open-source memory layer for people who use AI agents across sessions or switch between coding tools. It keeps preferences, project decisions and other saved context in local Markdown files that connected agents can share. The MIT-licensed Python tool works with Claude Code, Codex, Cursor, GitHub Copilot and other agents through MCP and editor integrations.

Agents turn notes, transcripts, PDFs and screenshots into a wiki with links back to the original sources. Durable memories require your approval. Session hooks supply context when an agent starts and propose memories when it finishes, so an automatic capture doesn't become an accepted fact. You can inspect the files in any editor or open the wiki in Obsidian.

Recall returns compact context with source references and confidence labels. The default memory layer doesn't need an LLM, API or vector database. Optional semantic matching uses a local embedding model and works offline after initial setup. Review dates bring stale memories back for checking, expiry dates remove them from default recall, and historical queries retrieve what was recorded as true at an earlier date.

Memory stays on your machine unless you choose to sync reviewed files through a Git remote you control. Sync checks for secrets and excludes private captures; teams can share approved memory through the same approach. A local viewer lets you browse the workspace, while LinkBar on macOS adds recall and memory review from the menu bar. Local retrieval records show which memories agents read and when, without recording your questions.

Similar to Link