Keep Every Claude Code Conversation

If you build with Claude Code, the most valuable thing you produce is often not the code. It is the conversation that produced it: why one design won, what was tried and thrown away, which constraint ruled out the obvious answer, what the next session needs to know.

By default, Claude Code keeps each conversation for 30 days and then deletes it, without a message. The setting is cleanupPeriodDays, and most people never touch it. Long before the 30 days are up, compaction has already removed the full tool outputs and intermediate reasoning from what Claude can see in a long session.

Every developer working with AI is generating a record of their own thinking, and almost none of them are keeping it.

The summary is not the record

Everyone who works with Claude Code learns the workarounds: CLAUDE.md, memory files, hand-off notes, pasting a summary into the next chat. Keep using them. They are curated guidance that Claude loads every session.

But each one is a summary that someone chose to write. It is partial by design, and it goes stale unless someone maintains it. When you need to know why a decision was made three months ago, the summary usually says what was decided, and the reasoning is gone.

An archive is different in kind. It is what actually happened, exactly, including the parts nobody thought were important at the time.

Rented models, owned context

There is a longer argument here, and it is the one we care most about. Models improve every few months, and each generation reads more carefully than the last. A conversation you had today will be read better by next year's model than by today's.

So the raw record appreciates. Summaries are readings of it made by one model era; they can be regenerated. The exact conversation cannot. If it was deleted, no future model can recover the thinking it held. If it was kept, unedited and on your own machine, it becomes a primary source for your own work: what you knew, when, and why you chose what you chose.

That is also why it should be yours. A record held in someone else's cloud is a record you rent access to. A record on your own computer, in an open format, is one you own and can take anywhere.

What Rootz Archive does

Rootz Archive is a Claude Code plugin. It is free for individual users. Once you accept its licence, it keeps:

• every Claude Code conversation, byte for byte, including what came before each compaction;
• sub-agent conversations;
• every saved version of every file Claude edits.

Everything stays on your own computer, readable only by you. Each saved version is fingerprinted (SHA-256) and re-checked daily, so you can tell if anything changed. The plugin makes no network connections, needs no account, and sends nothing to us.

Then you just ask Claude: “What did we decide about the login flow last month?” Claude searches the archive through a local MCP server, including sessions Claude Code has since cleaned up from its own history.

What it did for us

On one of our own laptops, over 30 days, Claude reached into its archive 60 times, in 17 of 265 sessions, on 14 different days. In the same period, 55 older conversations (21,966 messages) had already been deleted from that laptop by the 30-day cleanup. We still have them only because a copy existed.

That is one laptop, not a study, and we say so. The dashboard in the plugin follows the same rule: it shows what is protected, what is verified and how far back you can search, and it says “not measured yet” for anything we cannot yet measure on your own work.

Start in two commands

claude plugin marketplace add rootz-global/rootz-archive
claude plugin install rootz-archive@rootz

Then open a new Claude Code session and type /rootz-archive:accept. Nothing is captured until you do. It needs Node.js 22.13 or newer. Type /rootz-archive:dashboard any time to see what it holds, and /rootz-archive:pause or /rootz-archive:exclude-project when you want it to stop.

What it is not, and what is new

It is not a backup service: it keeps a copy on the same computer, so keep your own backups. It captures Claude Code today; other AI tools can search the archive through MCP, and the site marks which setups we have tested. One update since launch: you can now share a session with a capture receipt (/rootz-archive:share). The share is signed with a key made on your own computer, so anyone can check it came from that machine unaltered. On its own, that shows which device signed it, not who stands behind the device; the receipt says “Authorized by: verified” only when you add on-chain evidence naming the device.

Origin starts with keeping the record. Everything else, including proving where an AI’s work came from, depends on it having been kept.