Remnic Relay: correct once, every agent learns

Agent teams have a memory problem that single agents do not. When five agents share context and one of them learned something wrong, the error does not stay put. It spreads through summaries, handoffs, and shared files until nobody can say where it started.

Remnic Relay is our answer, built during OpenAI Build Week 2026. It treats correction as a first-class team operation.

What Relay does

Relay adds four things on top of the Remnic memory engine:

1. It exposes belief conflicts. When two agents hold contradicting memories about the same fact, Relay surfaces the pair instead of letting recall pick a winner silently.
2. It binds failures to stale memory. When a mission step fails, Relay traces which memory the agent acted on, so the fix targets the cause rather than the symptom.
3. It requires human-approved correction. No agent rewrites team memory on its own authority. A correction is proposed, reviewed, and applied through the correction contract.
4. It proves cold-agent recovery. A fresh agent with no session history joins the team, recalls the corrected state, and completes the mission that previously failed.

That last one is the point. Correct once, and every agent learns, including the ones that were not running when the correction happened.

Why this matters beyond the demo

Most memory systems measure recall. Very few measure what happens after a user says “that is wrong.” Our position is that a memory layer that cannot take a correction is a liability, and a team memory layer that cannot propagate one is worse.

Relay ships in the Remnic repo with committed fixtures and a no-install verifier on Linux. The broader correction work it builds on is measured by MemCorrect, our correction benchmark.

If you run more than one agent, this is the part of Remnic to watch. Start with the self-hosting page to see the team setup, or install Remnic and try it on your own machine first.