How do I keep my Claude Code sessions after 30 days?

Claude Code deletes session transcripts after 30 days, because they can hold secrets. That also deletes the fix you found last month and the reason you chose one library over another. Personal mode keeps that history on your laptop and lets the agent search it, with the secrets labelled and held back.

It is free and open source under Apache-2.0. It is the same service, policy check and audit log we sell to agent vendors, with a policy made for one developer.

pip install -e .
inhold init --personal               # makes ~/.inhold, prints MCP commands
inhold import claude-code --repo .   # sessions, memory, CLAUDE.md
inhold import codex                  # ~/.codex/sessions

Then run the claude mcp add or codex mcp add line that init printed. --repo also brings in that repo's CLAUDE.md and AGENTS.md. Run import again whenever you like: only new turns are added.

We imported a two-turn synthetic session. The first turn pastes an AWS key (the example key from AWS's own docs). The second records the fix.

$ inhold import claude-code ./sessions
{"files": 1, "items": 2, "denied": 0}

coding agent, purpose dev, searches "Postgres client Dockerfile"
  [internal] assistant: Retried. The fix was pinning the Postgres client to 16 ...

coding agent, purpose dev, searches "AWS_SECRET_ACCESS_KEY"
  no results

you, purpose vault, search "AWS_SECRET_ACCESS_KEY"
  [SECRET] user: The staging deploy failed. Use AWS_SECRET_ACCESS_KEY=...

From our run on main 98cf618, 31 Oct 2026, through the Python service the MCP tools call. The turn with the key is labelled SECRET. The agent's purpose cannot read SECRET, so that turn never reaches it; the turn next to it still does. You can read it under the vault purpose. Every import, search and read writes an audit row.

• The boundary holds for what the agent asks Inhold for. An agent with a shell running as you can still read local files, including ~/.inhold and the raw transcripts.
• The MCP command init prints carries the agent token, so the token lands in your agent's config.
• Secrets are found by pattern. A password with no = or : after its name is not caught.
• Import reads only the local paths above. Tool calls and tool output are skipped.