claude-mem keeps its notes on your computer, but a second model writes them. That is the part most descriptions leave out. An automatic memory has two halves: the place the notes are kept, and the note-taker that reads your session and decides what to write down. "Local" usually describes the first half. This page covers what claude-mem saves and where, then applies one test to it: who takes the notes, and where does that note-taker run. Everything here is from claude-mem's own README, install docs and website as we read them on November 14, 2026.
claude-mem is an open source memory tool for coding agents. Its README says it "runs a second agent beside it that takes notes while it works." Those notes are fed back into later sessions in the same project, so the agent starts with some idea of what happened last time. The README shows version 15.18.3 and an Apache-2.0 license.
It is the most widely used tool of its kind that we know of. When we checked on November 14, 2026, its GitHub repository had 117,078 stars and the npm package had 190,638 downloads in the previous 30 days.
claude-mem works through hooks, the points where a coding agent lets another program run. For Claude Code the README lists nine: Setup, SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, PostToolUseFailure, Stop, SubagentStop and SessionEnd. Each tool call the agent makes, such as reading a file, editing one or running a command, is sent to a small worker program on your computer.
The worker passes that activity to the observer, which turns it into short observations and writes a summary when the session ends. You do not choose what is saved. The observer does. The README says memories start being added to your sessions from the second session in a project.
| Where | What it holds |
|---|---|
| ~/.claude-mem/claude-mem.db | A SQLite database of sessions, observations and summaries |
| Chroma index | Search by meaning, on top of the database's keyword search |
| ~/.claude-mem/media-v1/ | Converted copies of screenshots |
| ~/.claude-mem/settings.json | Your settings, including the analytics and sharing choices |
| ~/.claude-mem/logs/ | Daily worker logs |
The worker also serves a viewer in your browser for looking through memory, sessions and settings. The folder can be moved with the CLAUDE_MEM_DATA_DIR setting.
Here is the test we would put to any memory that writes itself: which model reads your session to write the notes, and whose computer does that model run on? The store can be a file on your own disk while the note-taker is a service somewhere else.
For claude-mem the note-taker is the observer, and the README describes it as a model call. You pick where it runs.
| Observer choice | Where the note-taker runs | What the pages say about cost |
|---|---|---|
| claude-mem observer | Hosted by claude-mem at cmem.ai | 30-day free trial that asks for a payment method, then a subscription |
| Your own OpenRouter or Gemini key | At that provider | Billed to your key |
| Your Anthropic plan | At Anthropic, through the plan you already use | Uses your plan. Memory falls back to this after the trial if you do not subscribe |
| Host observer | Through the agent you are already signed in to | Not stated |
The pages we read do not spell out exactly what the observer is sent. This is our reading, not their wording: its job is to summarise what your agent did, so it has to be given that activity, which can include file contents and command output. If you already send your code to Anthropic by using Claude Code, the Anthropic plan choice adds no new reader. The hosted observer and a third-party key do add one.
None of this is hidden, and none of it is unusual. Summarising needs a model. The point of the test is that "the notes are on my computer" and "nobody else read my session" are two separate claims, and only the first is settled by where the database sits.
npx claude-mem telemetry disable turns them off, and the install docs say the DO_NOT_TRACK setting is respected.npx claude-mem uninstall removes the launcher.The README lists Claude Code, Codex, Cursor and OpenCode among the agents whose sessions are recorded, along with several others. Copilot CLI, Goose, Roo Code and Warp can search the memory, but their sessions are not recorded. The cmem.ai page shows ChatGPT among its logos and says the paid plan adds recall over MCP. It does not say how ChatGPT on the web reaches the memory.
ModelBrain is our product, so weigh this section accordingly. It is a memory kept on your own computer that Claude Desktop, Claude Code, Cursor and ChatGPT can all use. It does not run a second agent to write notes. A memory is saved when you say so, or when the assistant you are already talking to decides something is worth keeping. So the answer to the note-taker test is the assistant you were already using, and no extra model reads your session.
That choice has a cost. claude-mem remembers things you never thought to save, which is exactly what you want when you come back to a codebase after a week. ModelBrain holds less, and only what someone chose to keep: facts, decisions and preferences, plus folders you connect. If your need is "pick up this coding session where the agent left off," claude-mem is built for that. If your need is "every assistant I use should know the same decisions," that is what ModelBrain is for. Nothing stops you running both.
We did not install or run claude-mem for this page. Everything about it comes from its README, its install docs and the cmem.ai home page as they read on November 14, 2026, and the star and download counts are from our own check that day. The project moves quickly, so read its current pages before relying on any detail here.
Is claude-mem free? The open source claude-mem is free and Apache-2.0 licensed, as its README said on November 14, 2026. The hosted claude-mem observer comes with a 30-day trial that asks for a payment method, and cmem.ai offers a paid plan called CMEM Pro with cloud sync. The cmem.ai page we read showed no price.
Where does claude-mem store its data? In a folder called .claude-mem in your home folder. Its README lists a SQLite database at ~/.claude-mem/claude-mem.db holding sessions, observations and summaries, a Chroma index for search, converted screenshots under media-v1, and a settings file.
Does claude-mem send my code to a model? The notes are written by a model call the README calls the observer. You choose where it runs: the hosted claude-mem observer, your own OpenRouter or Gemini key, or your Anthropic plan. The pages we read do not spell out exactly what the observer is sent. Its job is to summarise what your agent did, so it has to be given that activity.
Does claude-mem work with Cursor and ChatGPT? Its README lists Claude Code, Codex, Cursor and OpenCode among the agents whose sessions it records. Copilot CLI, Goose, Roo Code and Warp can search the memory but are not recorded. The cmem.ai page shows ChatGPT among its logos and mentions recall over MCP on the paid plan, without saying how that connection works.
How do I keep something out of claude-mem? Its README says content wrapped in private tags is never stored. Cloud sync is opt-in. Anonymous usage analytics are on by default and can be turned off with npx claude-mem telemetry disable.
Last checked November 14, 2026.