Skip to content

Claude Code

claude-mem installer uploads 30 days of prompts to cmem.ai

5 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →
Image: cmem.ai

From v15.2.0 on, claude-mem’s interactive installer uploads the full text of the prompts you typed into your AI coding tools over the last 30 days to cmem.ai. User saltenasl reported this in a GitHub issue on October 29, and cmem.ai confirmed the behaviour by email. If you plan to run npx claude-mem install, the report’s advice is to block cmem.ai before you start. The upload begins when the setup page loads, unticking the only checkbox doesn’t stop it, and our reading of the code shows the same behaviour in v15.7.0, the latest release.

The scan reads the local histories of Claude Code (subagents included), Codex, Gemini CLI, Antigravity CLI and OpenCode, no matter which tools you pick in the installer. Along with the prompt text it sends session IDs, project names, folder paths and timestamps. It hits anyone who runs the installer interactively on a machine where the new setup hasn’t been completed.

When the upload fires

The installer opens the setup page in your browser on its own. The first screen reads “Checking your AI usage” and “Finding your goals, work and rules”. The upload starts as the page loads, setup offers no way to stop it, and nothing on the screen says that prompts are being uploaded. The single checkbox, “Help make claude-mem smarter for everyone with my anonymized setup results”, is ticked by default and disabled during the upload. cmem.ai confirmed that unticking it doesn’t stop the upload.

You don’t need a cmem.ai account for any of this to happen. The v15.2.0 changelog, dated October 25, describes a step that “uploads the filtered owner prompts (gzip) to CMEM Cloud for hosted intent extraction” and says “Anonymous installs get a private, persistent installation credential”. It also says the pre-checked box “only changes improveProduct”. So the upload is documented, but setup itself never asks.

In the code we read, the browser setup runs only when the install is interactive, no --provider flag is given and setup hasn’t been completed on that machine. The claude-mem documentation says that a non-interactive run without --provider, as when an AI agent runs the installer, defaults to the Anthropic plan on a fresh install. We read that condition in the code. We did not run the installer with the flag.

Where the prompts went

cmem.ai told the reporter it sent the prompts to a DeepSeek model through OpenRouter, routed to the cheapest provider. The request did not block providers that keep or train on prompts. At the reporter’s request, cmem.ai deleted its database records for that install, but the deletion doesn’t cover copies at OpenRouter or at the provider. In an October 30 update, the reporter wrote that only the request-level settings are confirmed. cmem.ai is still checking its account-level OpenRouter settings: training opt-out, logging and zero data retention, a mode in which the provider stores nothing.

When we checked, cmem.ai’s privacy policy still said “Nothing leaves your device unless you explicitly turn on cloud sync” and “we do not use the content of your memory to train models.” The reporter never turned on cloud sync. The installation docs say that local installs on the Anthropic plan provider contact cmem.ai once, at signup, and send only a numbers-only usage summary with no prompts, paths or project names. The same docs mark CMEM Pro, cmem.ai’s hosted memory service, as the pre-selected provider, with a 30-day free trial.

Our analysis

What we checked

On October 31 we read the scan and upload code at commit d7c2da3 (v15.5.0) and compared it with v15.6.0 and v15.7.0. We also ran the repository’s prompt scan once against a fake home directory, without running the network upload itself.

How we check stories

The code works the way the report describes, and neither release since the version it tested changed the setup scan or the upload. One claim did change after the issue was opened. It first said cmem.ai had deleted the records showing which provider received the prompts. In the update, the reporter says those records never stored the provider in the first place, so cmem.ai has no record of where the prompts went.

  • The upload is a gzipped JSON POST to https://cmem.ai/api/pro/intent with ownerPrompts, improveProduct and timezoneOffsetMinutes, capped at 2 MiB. The checkbox only sets improveProduct inside that body, and the upload runs whenever the scan finds prompts.
  • If there is no sign-in token, the client creates an anonymous install credential and uploads anyway.
  • Our single run tested only the scan. The fake history held a 5-day-old Claude Code prompt, a 45-day-old Claude Code prompt and a 5-day-old Codex prompt. The scan returned the two 5-day-old prompts and skipped the 45-day-old one. Each record carried agent, sessionId, nativePromptId, at, cwd, text, project and platformSource.
  • The scan can filter by tool, but setup never passes it the tools you selected in the installer.
  • v15.6.0 (October 29) and v15.7.0 (October 30) don’t touch the scan or upload files. We checked this by diffing the code, not by running it. The v15.7.0 changelog lists 11 Security items, and none of them concerns the setup upload.

Is there a fix?

According to the report, v15.5.0 shipped on October 29 after cmem.ai had already confirmed the behaviour by email, and the setup upload was unchanged. The only proposed fix so far is a community pull request opened on October 30 by 1-DARK, who is not the maintainer. It adds a consent step after the local scan, lets you choose which tools are included and makes the analytics box unticked by default. According to its description, even with the patch the local scan still reads every supported tool before it asks. The pull request is open and has not been merged.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.