Your agents are brilliant in the session and blind to everything around it — so you become the courier who pastes context back in. LUCI is persistent memory for AI agents: a 100% local, searchable record of your screen and meetings. Install the official skills so Claude Code, Codex, and Cursor can query it. Nothing uploaded. Nothing re-explained.
Agent memory is the persistent context an AI agent keeps across sessions — what it knows about you, your work, and what happened before, without being told again. Agents have short-term memory by default: the context window, which empties when the session ends. Everything else has to come from somewhere, and today that somewhere is usually you, pasting.
The industry's first answers are app-scoped: ChatGPT remembers things inside ChatGPT, Claude inside Claude, and OpenAI's Codex Chronicle distills your screen into memories only Codex can read. Useful — but your context ends up sharded across vendors: each app remembers its own slice, none can read another's, and none of them saw your screen or your meetings in the first place.
For the full definition and the taxonomy — short-term vs. long-term, distilled vs. record-based — see the glossary entry: What is agent memory?
Distilled memories (ChatGPT memory, Chronicle) summarize what seems durable and discard the rest. That works until the thing you need is the thing the summarizer dropped — the exact error text, the number quoted in a call, the setting you changed once. A distilled memory is an index of your work, not a record of it.
LUCI builds the record: everything that appears on your screen and everything said in your meetings, captured, transcribed, and indexed on your own machine. Agents get both kinds of recall — fast search over the whole history, and drill-down to the exact moment when it matters. The index-vs-record tradeoff is covered in depth in our comparison with Codex Chronicle.
Related: what that record looks like for developers, in computer history for AI agents.
An agent-usable memory of your work necessarily contains your messages, documents, credentials, and other people's words. Ship that to a vendor's cloud and you've built the world's most detailed dossier on yourself — on someone else's servers. The LUCI pipeline runs end to end on your device, encrypted at rest. Wi-Fi off, it keeps working.
Memory accumulated inside one assistant is leverage for that assistant's vendor — it can't follow you to a better tool. The official LUCI skills work in Claude Code, Cursor, Codex, and other agents, so the same memory follows you when you switch tools.
Agents reach LUCI through official skills, or through MCP if they cannot load skills. What an agent retrieves becomes part of its conversation — so you control which agents connect and what they fetch. The memory itself stays on this machine.
LUCI records your screen and meeting audio on your Mac or Windows machine, then uses on-device understanding to interpret what you saw and transcribes both mic and system audio — building LUCI Memory entirely on-device.
From Settings → Agent access, or from github.com/Memories-ai-labs/Luci-skills. The skills teach the agent when to search your memory and how to distill your day.
From then on, "what was I looking at this morning?" is a prompt, not a paste job. MCP is the fallback if an agent cannot load skills.
Agent memory is the persistent context an AI agent keeps across sessions — knowledge about you, your work, and prior interactions that survives after the context window empties. It's what turns an agent from a brilliant stranger into a colleague who was there yesterday.
Install LUCI, enable Screen Memory, and install the official skills from Settings → Agent access, or from github.com/Memories-ai-labs/Luci-skills. Claude Code, Cursor, and Codex can be written from the app. Gemini, Copilot, Windsurf, and others install from the repo. MCP is the fallback if an agent cannot load skills.
Official skills install into Claude Code, Cursor, and Codex from the app, and into Gemini, GitHub Copilot, Windsurf, and other agents from the Luci-skills repo. MCP is available if an agent cannot load skills.
No — those are app-scoped memories stored on the vendor's side, readable only inside that product, and blind to your screen and meetings. LUCI is a cross-agent memory of your actual work, stored on your machine, that any connected agent can search.
The memory itself stays on your machine — capture, understanding, and storage are all local. What an agent retrieves becomes part of that agent's conversation, so you choose which agents you connect and what you ask them to fetch.
Exact words on screen, fuzzy recall by meaning, captures narrowed by app and time, a whole range at once ("what was I doing between X and Y"), what was said across mic and system audio, and any single moment in full detail. Start with the official skills; MCP is the fallback.
100% local. Install official skills for Claude Code, Codex, and Cursor.
Free on Mac & Windows. Your memory stays on your device, encrypted at rest.