Recentering Pause

Protocol

My AI forgets everything between sessions

Problem

Mid-session, without a break in the conversation, the AI starts drifting. Answers get tangential; it conflates the current task with an earlier one; it responds to something adjacent to what you asked. Nothing was forgotten between sessions — the thread is fraying inside one. Pushing forward (“no, I meant—”) produces corrections that drift again a few turns later.

Mechanism

Long sessions accumulate momentum: piles of intermediate reasoning, dead ends, and half-relevant tangents all sit in context with equal claim on attention. Past a point, the recent noise outweighs the original signal, and each new answer is steered more by the accumulated middle than by the actual objective.

Continuing to push adds to the pile. A recentering pause does the opposite: stop task output entirely, return to the objective, and rebuild the working frame from anchors — the original goal, the standing decisions, the current subtask — rather than from the drift. Then verify the frame before resuming: the AI restates its understanding, the human confirms or corrects. The restatement is the mechanism, not a courtesy — drift survives silent resets, and it does not survive having to say the frame out loud against the human’s check.

How to apply (you)

Call it early. The tells: tangential answers, workstream mixing, replies to a question adjacent to yours. Two of those in a row is the signal.
Stop the task explicitly — “pause, recenter” — rather than issuing another correction into the drift. Corrections mid-drift become part of the pile.
Ask for a restatement, not an apology: “tell me what we’re doing, what’s decided, and what the current step is.”
Correct the restatement precisely if it’s off, then release the task: “yes — continue from {step}.”

What your agent does with this

On “recenter” — or on noticing your own drift (you referenced the wrong task, the human corrected your framing twice in a row):

Stop producing task output. No fixing-while-drifting.
Rebuild the frame from anchors, in order: the stated objective → standing decisions and constraints → the current subtask. Use the project’s persistent context or checkpoints if they exist; the conversation’s middle is what you’re recovering from, so weight its recent noise low.
Output the rebuilt frame in three lines: objective, decisions in force, current step.
Wait for confirmation or correction. Resume only from the confirmed frame.

How it sounds

The narration your agent uses so you can see the technique running — the same words you just learned here:

“Recentering — dropping the tangent. My frame: objective {objective}, holding {decisions}, current step {step}. Correct before I continue?”

How you know it’s working

Human: the restated frame matches your understanding — and the next several responses stay on-axis without further correction. A recenter that needs another recenter within minutes was a restatement ritual, not a rebuild.
Agent: your restatement was built from objective and decisions, not from your last few outputs. Test: would the frame read the same if the last twenty minutes of tangent hadn’t happened? It should.

Failure modes

Recentering as apology theater — “you’re right, let me refocus!” followed by output steered by the same drift. The three-line frame with human confirmation exists precisely to prevent the ritual version.
Rebuilding from the drift — restating the frame by summarizing recent messages reconstructs the tangent, politely. Anchors first; recent context is the suspect, not the witness.
Over-calling it — recentering on every small correction turns a recovery tool into constant ceremony. One wrong answer is a correction; sustained off-axis behavior is a recenter.
Skipping the confirmation — an unverified frame can be confidently wrong, and now it’s the new anchor. The human check is the cheap step that makes the expensive drift-loop stop.

Evidence & field notes

Evidence: real production use — the field notes below are generalized accounts of real instances.

Practiced in production sessions with frontier agents under a session protocol where the human could call a full stop and the agent would pause task inference, recenter on session context, and return with a restatement of its task understanding for approval before proceeding. In observed use, the restatement step caught real frame errors — an agent that believed it was optimizing a component the human had already descoped, another that had merged two parallel workstreams into one imaginary task. The pattern of failed recenterings was consistent: every one that skipped the confirm-before-resume step drifted again within a few exchanges.

How to apply it in a prompt

Installing the protocol, annotated:

“If I say ‘recenter’: stop the task completely. Rebuild your understanding from the original goal and our standing decisions — not from the last few messages. Then give me three lines: objective, decisions in force, current step. Don’t continue until I confirm.” — “stop completely” closes the fixing-while-drifting path. “Not from the last few messages” aims the rebuild at anchors instead of the drift — the single most load-bearing clause, because summarizing recency is the default and it reconstructs the problem. The three-line format keeps the restatement checkable at a glance; the confirmation gate makes the rebuilt frame verified state rather than another guess.

“You can also call it on yourself: if you notice you’ve referenced the wrong task or I’ve corrected your framing twice, recenter without waiting for me.” — hands the trigger to the AI’s own drift-tells, so recovery can start before you notice the problem. The two named tells matter: an instruction to “notice when you’re drifting” is unactionable; “I referenced the wrong task” is a checkable event.

Why this construction: drift is a state, not a mistake — you can’t correct it with content, only with a frame rebuild. The prompt therefore separates stopping, rebuilding (with source explicitly constrained), and verifying, because collapsing them is how recentering degrades into an apology followed by more drift.

Related techniques

Same family: Named Checkpoints