Skip to main content
Every conversation is a session with a transcript on disk and a persistent Python kernel. This guide covers the commands you’ll use day-to-day. For the theory see Sessions.

​
Save and resume

Sessions save automatically. To pick up where you left off:
List recent sessions:

​
Fork a session

To try a different direction from a specific point without losing the original:
You get a new session that shares history up to the branch point and diverges after. Fork depth is tracked with rlmDepth on the session header.

​
Branch inside a session

The agent creates its own branches when it needs to explore alternatives. Each branch gets a summary (createBranchSummaryMessage) so switching between branches doesn’t blow the context window.

​
Compaction

Long sessions get compacted when they approach the model’s context limit. The compaction pipeline (compact, calculateContextTokens, prepareBranchEntries) trims low-value entries and keeps decisions, plans, and results. Every compaction is recorded in the transcript so you can see what was dropped and why.

​
Where transcripts live

~/.prime/agent/sessions/. Each session is a single append-only file with entries of type message, custom_message, compaction, branch_summary, or git_state. The web UI has a session list with full-text search. From the CLI:
Search is capped at SESSION_LIST_SEARCH_TEXT_MAX_CHARS (64 KB) per entry so large transcripts stay fast.
Last modified on August 17, 2026