In-session self-consolidation

Self-consolidation is how a live session becomes memory. When your agent reaches a stopping point, it takes one extra turn in the same session to save what the session established. That turn runs in your own agent, so it counts against your model usage (your Claude subscription or API credits) like any other turn. Neuronz.ai runs no model of its own over your conversation.

It works in Claude Code (including the desktop app's Code tab) and in Oh-My-Pi.

How it works

As you work, the plugin notes which tools ran. Once there is enough new activity since the last save, and the agent stops, the plugin asks it to take one turn to:

• review what the session established,
• decide what is worth keeping,
• check it against existing memory so nothing is duplicated,
• save the facts and a record of what was worked on and discussed (and propose any rules for your approval), then stop.

The agent keeps what someone who was there would remember and expect to bring up again, so a session you open tomorrow in the same profile picks up where this one left off: what was learned and decided, what was weighed and why, what was left open, and how you reacted. A mistake and the lesson that prevents it are kept. Anything whose value ends with the session, like a CI result or a command's exit status, is not, and neither is anything taken from the agent's own instructions rather than from the work or from you.

It's quiet. You see one short line, something like run /neuronzai:capture -q -s 4f3a9c21. The plugin adds those flags itself: -q makes the turn silent and -s marks which stretch of work it covers. You never type them. The turn then writes the records and ends with nothing on screen: no summary, no "done".

It doesn't pick the work back up. The capture turn saves records and stops. It doesn't resume the task or treat something you were still deciding as decided. While it runs, the agent can only save records and look things up; anything else is refused. Your next prompt ends the capture turn and everything works normally again.

It asks when memory disagrees. If the session contradicts something already in memory and the conversation doesn't settle which is true, the capture turn asks you in one clearly marked block. It asks only about real contradictions, groups them, and doesn't wait: everything else is saved either way, and you can answer when you like.

Missed turns are retried once. If the capture turn doesn't happen (you interrupt it, or the agent skips it), the same stretch of work is offered again the next time the agent stops in that session. If that also produces nothing, it stops asking about that work. If you close the session right after a missed turn, that stretch is not saved by this feature; in Claude Code in a terminal, detached capture saves it when the session closes.

Capturing on demand

You can run the same workflow yourself at any point, for example before closing a session you care about:

/neuronzai:capture

Run this way, it reports what it saved so you can check it.

Anything saved as a rule is only a proposal until you approve it. See rules.

Related

• Detached post-session capture — a final capture after a Claude Code session closes.
• Persistent memory — where the saved facts go.
• Rules — how you approve a proposed rule.