Skip to main content
Back

OpenChronicle Gives AI Agents Durable Local Memory Without Vendor Lock-In

Einsia released OpenChronicle, an MIT-licensed local-first memory layer that captures your Mac's accessibility tree and feeds any tool-calling agent.

Repo, Full Access
  • Einsia released OpenChronicle, an MIT-licensed local-first memory layer for tool-capable LLM agents
  • Uses macOS accessibility trees as the primary capture signal instead of screenshot-plus-OCR pipelines
  • Stores memory as inspectable Markdown files backed by a local SQLite FTS5 index
  • Ships MCP integrations for Claude Code, Claude Desktop, Codex, and opencode out of the box
  • Model-agnostic: works with Ollama, LM Studio, OpenAI, Anthropic, or any LiteLLM provider
  • Currently v0.1.0 alpha, macOS 13+ only, with 2.8k stars and active contributor recruitment

A new open-source project called OpenChronicle just landed on GitHub, taking aim at one of the most persistent gaps in agent tooling: durable, local memory that survives across sessions and works with any model. Built by developer Einsia and a small group of contributors, the repo has already collected over 2,800 stars and pitches itself as an open, hackable alternative to closed screen-memory products.

The core promise: open-source, local-first memory for any tool-capable LLM agent. Think OpenAI Chronicle, but open, model-agnostic, inspectable, and hackable. It runs on your Mac, captures structured context from what you're doing, and turns it into persistent Markdown memory covering what you're working on, decisions you've made, tools you use, and which people or projects matter.

Accessibility trees over screenshots

The most interesting design decision is what OpenChronicle skips. Rather than a screenshot-plus-OCR pipeline like most screen-watching tools, it leans on macOS accessibility (AX) trees as its primary signal, reading structured text describing the focused element, visible text, URLs, and app state directly from the OS.

The tradeoffs the maintainers call out:

  • Cheaper processing. Structured text is far lighter than vision pipelines running over screenshots.