Import past conversations

Live capture starts when you install The Librarian. Everything you said to your agents before that is still on disk, in each harness’s own history. librarian import reads it on the machine where it lives and sends it through the same two steps live capture uses: memory intake (memories) and the journal.

librarian import --dry-run   # see what would be imported, send nothing
librarian import             # show the plan, ask, then send

Run it on each machine you’ve used agents on. It uses the same server address and token as the harness integrations (librarian config).

What gets imported

Each step only takes conversations that ended before that step first ran on your server. After that point live capture already has the conversation, so importing it again could only duplicate. The server keeps one cutoff per step:

Step	Its cutoff
Memories	When curator intake first ran
Journal	The start of your earliest journal entry

If a step has never run, every past conversation qualifies for it. Its cutoff is set the first time it runs, or the first time an import uses it.

Each cutoff is set once and never moves. Imports write journal entries dated in the past, and they start intake runs. If the cutoffs were worked out again each time, a second import, or an import from a second machine, would see a different rule. So you can import from as many machines as you like, in any order.

The server also keeps a ledger of conversation ids, written whenever capture finishes a step, whether the conversation arrived live or by import. A conversation already in the ledger is never imported twice. That covers re-running the importer, a resumed import, and a history folder synced between machines.

Some conversations are never imported:

• conversations that ended in the last day, because live capture still covers them;
• anything older than the server’s maximum age;
• subagent conversations, tool calls and output, and the model’s reasoning, exactly as with live capture;
• the recap a harness writes when it compacts a long conversation;
• anything inside a private-mode span.

The server’s limits

Every imported conversation costs the same model calls as a live one. Two limits on Settings → Curator → Intake → Import keep an import from running up a bill or flooding the queue:

• Maximum age (default 90 days): older conversations are never imported.
• Allowance (default 200 conversations): each imported conversation uses one. At 0, imports stop until you raise it. The importer says so and tells you how many were left.

The server enforces both limits, so a modified or outdated importer can’t get past them. Imported conversations are processed a few per sweep, after live capture, so an import never slows down today’s capture. They appear in your memories and journal over the following hours.

Options

Flag	What it does
--only memories / --only journal	Run just one step. If you’ve used The Librarian for a while, --only journal fills in the journal for your earlier conversations.
--harness claude,pi	Only read these harnesses. The default is all five.
--dry-run	Show the server’s plan and send nothing. Only conversation ids and end times go to the server; no text.
--yes	Don’t ask before sending.

If both intake and the journal are off on the server, the importer says so and stops without reading anything. If one is off, it warns you and imports into the other.

Where each harness keeps its history

Harness	Location	How long it’s kept
Claude Code	~/.claude/projects/	30 days by default (cleanupPeriodDays)
Codex	~/.codex/sessions/	No automatic deletion
OpenCode	~/.local/share/opencode/opencode.db	No automatic deletion
Pi	~/.pi/agent/sessions/	No automatic deletion
Hermes	~/.hermes/state.db	Kept unless sessions.auto_prune is on

Claude Code deletes conversations older than 30 days by default. You can raise cleanupPeriodDays in Claude Code’s settings if you want to keep more, at the cost of disk space. The Librarian doesn’t change it for you.

Reading OpenCode and Hermes histories needs Node.js 22.13 or later, for its built-in SQLite support.