Short-term vs. long-term agent memory
Every AI agent has short-term memory by default: the context window, which holds the current conversation and whatever you paste into it. It vanishes when the session ends. Long-term agent memory is what survives — facts, preferences, project context, and history that the agent can draw on next week without being told again.
Long-term memory is usually built one of two ways. The first is distillation: the agent (or a background process) writes down summaries of what seems durable — your preferences, your stack, recurring workflows. ChatGPT’s memory and OpenAI’s Codex Chronicle work this way. The second is retrieval over a record: keep a full, searchable log of the source material and let the agent query it at answer time. The two differ in one crucial way — what happens to details the summarizer didn’t think mattered.
The index-vs-record problem
A distilled memory is an index of your work, not a record of it. Anything judged one-off is dropped at write time and gone for good — and it is precisely the one-off details (the exact error text, the number quoted in a call, the setting you changed once) that you end up needing. A record-based memory keeps the source, so the agent can always drill from "you were debugging auth last Tuesday" down to the exact moment on screen.
The strongest setups combine both: distilled context for cheap, fast recall, backed by a searchable record for everything the distillation dropped.
Why agent memory should be yours, not the vendor’s
Most agent memory today is siloed and cloud-bound: ChatGPT’s memory lives in ChatGPT and only ChatGPT can read it; Chronicle’s memories are built on OpenAI’s servers and readable only by Codex. That means your accumulated context — arguably the most valuable byproduct of working with AI — ends up locked to the vendor that captured it. A memory that lives on your own machine and speaks an open protocol (MCP) inverts this: every agent you use gets the same memory, and none of them owns it.
How LUCI provides agent memory
LUCI builds the record layer: your whole day on the computer — meetings, code, email, documents, browsing — captured on your own Mac or Windows machine and stored encrypted at rest. The Agent Bridge gives that memory to whichever agent you use: works with Claude, Cursor, and Codex — LUCI connects itself, nothing to set up. Your agents can search what you saw, recall what was said, and reconstruct what you were doing, across sessions and across vendors, while the memory itself stays on your device.
Frequently asked questions
What’s the difference between agent memory and the context window?
The context window is working memory: whatever fits in the current session, gone when it ends. Agent memory is persistent: it survives across sessions and doesn’t depend on you re-pasting context. A good memory system feeds the right slices of long-term memory into the context window at the moment they’re needed.
Is agent memory just RAG?
Retrieval-augmented generation is the mechanism most record-based memories use: store the history, retrieve relevant pieces at answer time. Agent memory is the broader capability — it also covers distilled profiles and preferences. The practical question isn’t the acronym; it’s what gets stored, where it’s stored, and which agents can read it.
Don’t ChatGPT and Claude already have memory?
They have app-scoped memory: ChatGPT remembers things about you inside ChatGPT, Claude inside Claude. Neither can remember what happened on your screen or in your meetings, and neither’s memory is readable by the other. Cross-agent memory requires a store the agents share — which is what LUCI provides, on your own device: one memory of your whole day, readable by every agent you use.
How do I give my AI agent memory of my actual work?
Run a local memory that captures the work itself — meetings, code, email, documents, browsing — and let your agents read it. LUCI does this for you: it remembers your whole day on the computer, and the Agent Bridge connects itself — works with Claude, Cursor, and Codex, nothing to set up. They can then search what you saw, what was said, and what you were doing, with the memory stored on your own machine.
Is it safe to give an agent access to my memory?
Two separate questions. Storage: your memory lives on your device, encrypted at rest, and the Agent Bridge lets agents read only what they ask for. Retrieval: whatever an agent retrieves becomes part of that agent’s conversation, and cloud agents send their conversations to their vendor — so you stay in control of which agents you connect and what you ask them to fetch.