Imagine you and a few friends co-found a company. In the middle of the office sits one giant notebook—your company's "single source of truth."
The rule is simple: whatever anyone thinks of or discovers, they write it down directly.
On day one, friend A notes: "Client Wang prefers green tea." Fine. On day two, friend B, half asleep, scribbles: "The office printer runs on burning firewood." On day three, new colleague C opens the notebook, believes it completely, and actually carries a bundle of firewood 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 "read and forget," but instead have persistent memory and collaborate across many turns, every hallucination they produce can end up written in the public ledger, becoming the whole team's "institutional consensus."
A 2026 arXiv paper (《Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems》) makes a profound point: AI memory management can't rely on databases (retrieval) alone—it requires governance.
"Artificial Selection" in AI Memory
Author Diego F. Cuadros borrows a term from biology: Artificial Selection.
Darwin taught us that nature selects which genes survive through survival of the fittest. In the world of AI memory, we need to deliberately build a "filtering mechanism" that decides which information enters the "history books" and which must stay in "private diaries."
Let's break down this tiered memory governance system in Feynman style:
1. Private Diary (Agent-Local Memory)
Each AI agent has its own scratchpad. Its wild ideas, intermediate reasoning, and even occasional nonsense should be locked in its private space first. This is "nothing gets published without review."2. Institutional Truth (Shared Institutional Memory)
Only verified information of real value to the team may enter the "public ledger." The paper proposes this requires a set of "Selection Regimes":- Constitutional mode: filter information against a set of predefined hard rules.
- Automated review mode: let a separate "supervisor AI" audit information for accuracy and contradictions.
- Human review mode: at the most critical points, humans decide which memory becomes "permanent truth."
3. Archive
Every memory change must be traceable. If one day we discover "the printer burns firewood" was a mistake, we need to flip through old newspapers and find out who wrote it, when, and based on what rationale.Why Does This Matter?
In the past, we worried about AI making mistakes. In the future, we'll worry more about the "institutionalization of errors."
If an AI team handling legal documents has one member write an incorrect legal clause into the shared knowledge base, and every subsequent member uses it as their baseline, the entire system's legal advice collapses collectively.
The paper's core argument: we must upgrade AI memory systems from a simple "storage warehouse" into a rigorous "parliament."
Every permanently remembered piece of information must undergo a "survival contest" 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 bias 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 minding your 'public notebook'?"