Long-term memory for coding agents
Quit one agent.
Keep the thread.
Stop mid-task in one CLI, open another in the same directory, and carry over where you left off, what failed, and what is still open.
A protocol, not a convention. The project describes handoffs as typed, owned, and claimed exactly once, rather than a note left lying around.
One project, two sessions
Baton with the first agent
Session one
Agent A
holding the baton
Session two
Agent B
not started
claimed exactly once
- Where you left off
- The task as it actually stood.
- What failed
- The approach already tried and abandoned.
- What is open
- Questions nobody has answered yet.
Agent names are illustrative; the project lists over twenty supported harnesses.
The default path uses zero LLM calls: capture, search and handoffs work with no API key.