Don't Let AI Hallucinations Become 'History Books': Governing Collective Memory in LLM Multi-Agent Systems

Imagine you and a few friends start a company together, and in the middle of the office sits one giant notebook — the company's "single source of truth."

The rules are simple: whatever anyone thinks of or discovers, they write directly into the notebook.

Day one, friend A writes: "Client Mr. Wang likes green tea." Fine. Day two, friend B, half-asleep, scribbles: "The company printer runs on burning firewood." By day three, a new hire C opens the notebook, believes it completely, and actually hauls a bundle of firewood over to stuff into the printer...

This is the terrifying ghost currently haunting multi-agent systems: Memory Corruption.

When AI agents are no longer ephemeral chat boxes that "burn after reading," but instead possess persistent memory and collaborate across multiple turns, every "hallucination" they produce can be recorded on the public ledger and become "institutional consensus" for the entire team.

A 2026 arXiv paper ("Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems") makes a profound argument: AI memory management cannot rely on databases (retrieval) alone — it must rely on governance.

"Artificial Selection" in AI Memory

The author, Diego F. Cuadros, borrows a term from biology: Artificial Selection.

Darwin showed that nature selects which genes survive through survival of the fittest. In the world of AI memory, we likewise need to deliberately build a "screening mechanism" that decides which information gets written into the "history book" and which must stay in the "private diary."

Let's break down this tiered memory governance system the Feynman way:

1. Private Diary (Agent-Local Memory)

Every AI agent keeps its own scratchpad. Its wild ideas, intermediate reasoning steps, and even occasional nonsense should first be locked in its private space. Call it: "no publication without review."

2. Institutional Truth (Shared Institutional Memory)

Only information that has been verified and genuinely benefits the team may enter the "public ledger." The paper proposes "Selection Regimes" for this:

• Constitutional mode: filtering information against a set of predefined hard rules.
• Automated review mode: a separate "supervisor AI" audits whether information is accurate and free of contradictions.
• Human review mode: at the most critical junctures, humans decide which memory becomes "permanent truth."

3. The Archive

All memory changes must be traceable. If one day we discover that "the printer runs on firewood" was an error, we should be able to — like flipping through old newspapers — find out who wrote it, when, and on what grounds.

Why Does This Matter?

In the past, we worried about AI making mistakes.

In the future, we should worry more about the institutionalization of errors.

If one member of an AI team handling legal documents writes an incorrect statute into the "shared knowledge base," and every subsequent member treats it as the baseline, then the entire system's legal advice will collectively fail.

The paper's core thesis: we must upgrade AI memory systems from a simple "storage warehouse" into a rigorous "deliberative parliament."

Every piece of information permanently remembered must go through a "survival competition" and a compliance review.

To summarize:

We are evolving from "feeding AI data" to "teaching AI how to manage its own history."

Good memory is the foundation of intelligence; ungoverned memory is just a wastebasket full of biases and hallucinations. Future AI collaboration may be decided not by whose CPU computes faster, but by whose "historian" is more rigorous and whose "institutional memory" is cleaner.

The next time you see a swarm of AI agents busily working together, remember to ask: "Hey — who's managing your 'public notebook'?"