LUCI Desktop Review: Local Memory for AI Agents

What LUCI Desktop Actually Records

The capture is broad rather than app-specific. It watches the screen and listens to calls, then files what it finds against a timeline you can browse afterwards. Nothing is scoped to a single application, so the record covers the Slack thread, the meeting and the browser tab you had open while they were happening.

Call audio is transcribed on the device. The useful part is the link between the two: a sentence said out loud is stored against the screen state at that moment, so the reason behind a change is usually a quote rather than an inference from a diff.

A distilled summary keeps what a summariser judged durable and drops the rest. That is the layer an agent reads first, which makes answers fast. When the summary is not enough, the raw transcript and the screen history behind it are still on disk.

How Claude Code, Cursor and Codex Read the Memory

Official skills are installed for Claude Code, Cursor and Codex. The agent decides when to search the memory, so a question like why a particular architecture was chosen resolves without you pasting anything into the chat.

Gemini reads the same memory, which matters if you move between models. The record outlives any single tool, so switching assistants does not mean re-explaining your week.

There is also an MCP path for agents that do not use skills. It binds to loopback only and issues a token per install, so it is not reachable from the network by default.

What an Agent Gets Back When It Searches

A search returns two layers. The first is a distilled summary of what the day contained, and that is what an agent reads by default because it is small and quick. The second is the raw material behind it: the transcript of a call and the screen history at the moment it was made, still on disk and still indexed on the device.

The timeline is browsable by hand as well, in hour and day views, which is useful when the summary has thrown away the exact wording you needed. A summary is good at intent and bad at literals; an error string, a magic number or a quoted constraint is the sort of thing that only survives in the transcript underneath.

Because capture reads the screen and calls directly instead of connecting to named applications, every program you open is already supported. That includes internal tools and private documents that no integration would ever cover, which is usually where the context an agent is missing actually lives.

Privacy: Where the Data Lives and What Gets Redacted

Everything runs on your own machine. Recordings, transcripts and the memory index are processed and stored locally, and the company states they are not shared and are not used to train models. Understanding also happens on device, using local models, so the pipeline does not depend on a remote inference service.

Anything resembling a token or a credential is removed before the record reaches the disk. The store itself sits encrypted at rest on your machine, and the retention window runs from roughly a week to a year.

The honest caveat is about capture, not storage. Because it watches the whole screen rather than named apps, it will record whatever passes across it, including work that belongs to someone else. An opt-out per application is available, and password managers and banking apps are excluded by default, but on a shared machine the default is still a broad one.

LUCI Desktop Pricing: One Plan, No Upsell

There is a single free plan and nothing to buy. No card, no trial countdown and no usage cap on how much you record. There is no paid tier to compare against.

Because the app runs on your own hardware, the cost of the local summarisation falls on your machine rather than on a token bill. LUCI uses the NPU or GPU where one is available.

Pros and Cons

On the plus side it is free, genuinely local, and it connects to the coding agents people already use without asking for an API key. Meeting transcripts tied to screen context solve a real problem that a repository cannot.

On the minus side it is Mac and Windows only, and it adds nothing on its own. The value shows up when an agent asks it a question, which means you need Claude Code, Cursor or Codex in the loop as well. It is also a young product, so the troubleshooting trail is thinner than for tools that have been around for years.

Pricing & Plans

Free for Mac and Windows. No paid tier, no trial and no usage cap.

Best For

Recommended use cases and scenarios where LUCI Desktop shines.

Frequently Asked Questions

Common questions about LUCI Desktop, answered.