Skip to content

Privacy

  • Local. One SQLite file in your user folder. No telemetry, no cloud, no server, no account.
  • No API key. Facts are written by the agent’s own CLI under your existing login.
  • A request stays with its vendor. A Claude Code request is never sent to OpenAI, and the other way round.
  • Secrets are removed before anything is stored.
  • Not encrypted. Anyone who can read your home folder can read it.
StoredFor
your requests and the agent’s answers (cleaned, at most 16,000 characters each)search, the briefing, facts
tool calls: which tool, the command or file, a cut version of the output, whether it failedfix hints, file history, recipes
facts, decisions, rules, lessons written from themreminders
the notes Claude Code and Codex keep themselvesshown to the other agents

Tool output and inputs are trimmed after 30 days (retention.payload_days); the requests, answers and memories stay until you delete them.

  • anything between <private> and </private>, in a prompt, a file, or tool output;
  • secrets in known formats (API keys from Anthropic, OpenAI, GitHub, GitLab, Slack, Stripe, Google, AWS, npm, PyPI, Hugging Face; JWTs, bearer tokens, private keys, passwords in URLs, command lines, headers and assignments);
  • the content of files matching exclude_globs (.env, *.pem, *.key, id_rsa*, secrets/**, … and your own); only that such a file was touched;
  • web pages and third-party MCP output never go into the text the fact model reads;
  • agent-mem’s own processes and model calls;
  • anything in projects that are excluded, opted out or paused.

The full list of patterns is on Capture & Privacy.

Want to keep outDo
one piece of textwrap it: <private>the staging key is …</private>
certain filesadd globs to exclude_globs in config.json or "exclude" in the project’s .agent-mem.json
a whole project{"capture": false} in its .agent-mem.json, or its path in excluded_projects
everything for a whileagent-mem pause --for 2h (or until agent-mem resume)
one agent processstart it with AGENT_MEM_DISABLE=1
model calls"observe": {"enabled": false}: requests are still recorded and searchable, no facts
the agents’ own notes"native_notes": false
TrafficWhen
model downloads (search model about 220 MB, spaCy model about 13 MB)setup --write, models install, after a model change
your agent’s CLI talking to its own providerwhen facts are written, under your login, with the request’s text
nothing elsemem0’s telemetry is switched off and a test checks it