One note: treat this as a design concern, not a request for a specific implementation. You will obviously have a much better sense of what fits CORAL’s architecture.
Consider the long-run risk that shared memory can slowly make agents think more alike.
Picture one agent writing a note that is only partly true, or finding a strategy that improves the score in one narrow case. Let later agents reuse that note, copy the same strategy, build on the same assumption, and gradually stop exploring other directions. At that point, the system is still “collaborating,” but it may also be narrowing itself too early.
Add hallucination and memory poisoning to that risk.
Imagine an agent misreading an eval result, hallucinating a causal explanation, or writing a confident but wrong note into shared memory. Once that note is available to future agents, it can stop being “one agent’s mistake” and become part of the group’s working context. Later agents may cite it, reinforce it, build skills around it, or avoid useful paths because of it.
So the concern is not only bad memory in the simple sense. The concern is that shared memory can amplify errors over time:
- hallucinated observations becoming reusable context;
- weak assumptions becoming group doctrine;
- stale notes becoming invisible constraints;
- one-off skills becoming default behavior;
- agents optimizing the visible evaluator instead of the broader task;
- poisoned memory causing future agents to inherit the same blind spots;
- fast convergence, but toward a local optimum or false belief.
Separate this from ordinary agent mistakes. A single-agent hallucination is often local. A shared-memory hallucination can become contagious.
Add an anti-drift layer not to make CORAL less autonomous, but to preserve useful diversity and memory hygiene over longer runs. Use it to separate “interesting hypothesis” from “validated knowledge.” Let it retire stale skills, reduce self-reinforcing assumptions, and make it easier to see when a run started converging in the wrong direction.
So my suggestion is less “please implement this exact structure” and more:
Consider how CORAL can keep shared knowledge from becoming too self-reinforcing, especially as runs get longer and agent-generated memory becomes richer.
Keep the benefits of agent collaboration, while reducing the chance that the whole agent society gradually inherits the same hallucination, poisoned memory, or mistaken assumption.
One note: treat this as a design concern, not a request for a specific implementation. You will obviously have a much better sense of what fits CORAL’s architecture.
Consider the long-run risk that shared memory can slowly make agents think more alike.
Picture one agent writing a note that is only partly true, or finding a strategy that improves the score in one narrow case. Let later agents reuse that note, copy the same strategy, build on the same assumption, and gradually stop exploring other directions. At that point, the system is still “collaborating,” but it may also be narrowing itself too early.
Add hallucination and memory poisoning to that risk.
Imagine an agent misreading an eval result, hallucinating a causal explanation, or writing a confident but wrong note into shared memory. Once that note is available to future agents, it can stop being “one agent’s mistake” and become part of the group’s working context. Later agents may cite it, reinforce it, build skills around it, or avoid useful paths because of it.
So the concern is not only bad memory in the simple sense. The concern is that shared memory can amplify errors over time:
Separate this from ordinary agent mistakes. A single-agent hallucination is often local. A shared-memory hallucination can become contagious.
Add an anti-drift layer not to make CORAL less autonomous, but to preserve useful diversity and memory hygiene over longer runs. Use it to separate “interesting hypothesis” from “validated knowledge.” Let it retire stale skills, reduce self-reinforcing assumptions, and make it easier to see when a run started converging in the wrong direction.
So my suggestion is less “please implement this exact structure” and more:
Keep the benefits of agent collaboration, while reducing the chance that the whole agent society gradually inherits the same hallucination, poisoned memory, or mistaken assumption.