Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems

This paper frames the challenge of selecting persistent memories in LLM-based multi-agent systems as "governed collaborative memory," proposing a design agenda that treats memory governance as an artificial selection regime to ensure epistemic quality, provenance fidelity, and traceable institutional state rather than relying solely on retrieval accuracy.

The Gist

Imagine a team of AI assistants working together on a long-term project. In the past, these AIs were like strangers meeting for a single coffee chat: they talked, gave advice, and then forgot everything once the meeting ended. They had no "memory" of who they were or what they learned.

But now, these AIs are getting persistent memory. They can remember lessons from yesterday, store rules for tomorrow, and pass knowledge to their teammates. This is great, but it creates a new problem: Who gets to decide what becomes part of the team's permanent history?

If an AI makes a mistake, writes a funny but wrong story, or learns a bad habit, should that become a permanent rule for the whole team? Or should it stay private?

This paper argues that we need a "Governed Collaborative Memory" system. Think of it not just as a filing cabinet, but as a selection process—like how a museum curator decides which artifacts go on display and which stay in the basement.

Here is the breakdown of their ideas using simple analogies:

1. The Problem: The "Wild West" of Memory

Without rules, an AI might just save everything it thinks is interesting.

• The Analogy: Imagine a student who writes down every thought they have in a diary, including typos, daydreams, and false facts. If they later read that diary to decide what to do, they might act on a lie they wrote down by accident.
• The Risk: In AI terms, this is "ungoverned persistence." A false memory gets saved, reloaded, and repeated until it becomes a permanent, unchangeable "fact" for the whole system.

2. The Solution: Four Different "Memory Layers"

The authors suggest we shouldn't treat all memory the same. Instead, we should organize it into four distinct "rooms" in the house, each with different rules for what gets in:

• Room 1: The Personal Locker (Agent-Local Memory)

  • What it is: Private notes specific to one AI's role.
  • The Analogy: A chef's personal recipe book or a mechanic's specific tool preferences.
  • Why: If we force the chef and the mechanic to share the exact same notes, the chef might start fixing cars and the mechanic might start cooking. We need to keep their unique "identities" separate so they stay good at their specific jobs.

• Room 2: The Town Hall (Shared Institutional Memory)

  • What it is: The official, permanent rules and lessons for the whole team.
  • The Analogy: The city's official laws or the company's handbook.
  • The Rule: Nothing goes here unless it passes a strict "governance" check. It's not enough for an AI to just think it's a good idea; it needs proof and approval.

• Room 3: The Archive (Archive Memory)

  • What it is: Old history, research, and background info.
  • The Analogy: A library's basement or a museum's storage vault.
  • The Rule: You can look at these items, but they aren't active rules. We don't need to vote on every old newspaper clipping before someone reads it, but we must know where it came from.

• Room 4: The Whiteboard (Project-Continuity Memory)

  • What it is: Temporary notes for the current task.
  • The Analogy: A sticky note on a desk for today's meeting.
  • The Rule: This is erased or moved when the project is done. It shouldn't accidentally get mixed up with the permanent laws in the Town Hall.

3. How the "Selection" Works

The paper compares different ways to decide what gets into the "Town Hall" (Shared Memory):

• The "Let it All In" Approach (Ungoverned): Fast, but dangerous. Falsehoods become permanent facts.
• The "Test Score" Approach (Automatic): An AI checks if a memory improves a math score or speed. Good for numbers, but bad for things like "honesty" or "fairness."
• The "Rulebook" Approach (Constitutional): The AI follows a set of human-written rules (like "don't lie"). It's scalable but might miss nuances.
• The "Human Judge" Approach (Human-Ratified Artificial Selection): A human (or a human-led process) looks at the candidate memory and says, "Yes, this is true and important; let's make it official."
  • Why this matters: Humans are better at judging things that can't be measured by a score, like "did this AI sound trustworthy?" or "does this fit our team's values?"

4. What the Evidence Shows

The authors tested this idea in one real-world AI system. They found:

• Mistakes happen: Even with rules, an AI can still make a fake story.
• The system learns: Instead of just deleting the mistake, the system recorded why it was a mistake and created a new rule to prevent it next time.
• Identities stay safe: New AI team members could join and learn the team's rules without losing their own unique personalities.
• Transparency: The system kept a "paper trail" showing which memories were rejected, which were revised, and which were approved. You could see the history of the decision, not just the final result.

The Big Takeaway

The paper isn't saying "Humans must check every single memory." Instead, it says: We need to be intentional about how we select memories.

We need to ask:

1. What are we saving? (A fact? A feeling? A rule?)
2. Who decides it's good enough to be permanent? (A test? A rulebook? A human?)
3. How do we keep the AI's unique personality separate from the group's shared knowledge?

If we don't answer these questions, we risk building AI teams that are efficient but prone to repeating their own lies, losing their unique skills, or becoming a confused, identical blob of data. The goal is to make memory inspectable, correctable, and honest.

Technical Summary

Technical Summary: Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems

Problem Statement
As Large Language Model (LLM)-based multi-agent systems evolve from stateless participants in isolated interactions to state-bearing components with persistent memory, a critical design gap has emerged. While existing research addresses retrieval accuracy, access control, and memory consistency, it often overlooks the fundamental question of selection: which candidate memories should transition from private, transient data into shared, durable, and behavior-shaping institutional state?

Without explicit governance, persistent memory risks becoming a vector for "heritable" errors, where plausible but false summaries are reloaded and propagated across sessions, agents, and system versions. The core problem is not merely who can read a memory, but who or what decides that a candidate memory should become part of the system's shared ontology. This paper frames this challenge as governed collaborative memory, arguing that memory governance functions as a selection regime determining which memory variants persist, remain private, are rejected, or are superseded.

Methodology and Framework
The paper proposes a conceptual framework and a layered architectural design rather than a controlled experimental comparison. The methodology involves:

1. Defining Selection Regimes: The authors categorize memory governance into four distinct regimes, emphasizing that these are design choices based on target properties rather than a hierarchy of superiority:

   • Ungoverned Persistence: Memories enter durable storage without deliberate evaluation. This offers low friction but risks the persistence of falsehoods and hallucinations.
   • Constitutional/Hybrid Selection: Human-authored principles or constraints are enforced by automated processes (e.g., RLHF, Constitutional AI). This scales well but may miss traits omitted from the principles or succumb to reward hacking.
   • Automatic Metric-Based Selection: Persistence is determined by benchmarks, tests, or measurable rewards. This is effective for measurable performance but fails to capture judgment quality, epistemic honesty, or institutional coherence.
   • Human-Ratified Artificial Selection: An operator or governance process explicitly decides persistence. This is slower and less scalable but is necessary for selecting qualities that lack clean metrics, such as role preservation and institutional coherence.

2. Proposed Layered Architecture: To implement these regimes without collapsing heterogeneous agent identities, the authors propose separating memory into four distinct layers with different governance burdens:

   • Agent-Local Memory: Private by default; preserves role-specific heuristics, constraints, and failure modes to prevent identity collapse.
   • Shared Institutional Memory: A durable store for cross-agent events, principles, and lessons. This layer carries the heaviest governance burden, requiring evidence and explicit ratification.
   • Archive Memory: Long-horizon historical artifacts and background knowledge. These are retrievable but distinct from active institutional memory to prevent governance overload.
   • Project-Continuity Memory: Active task states and handoff contexts, kept separate to avoid scope contamination between transient work and durable knowledge.

3. Mechanism of Governance: The framework relies on provenance and version lineage. Selection is not just about the final accepted record but making the process inspectable. This includes tracking who drafted a memory, what evidence supported it, who ratified it, and what versions were rejected or superseded. The architecture adopts a "supersede-not-erase" semantics, ensuring that corrections remain traceable as institutional knowledge.

Evidence and Results
The paper supports its design claims with documented traces from a single running LLM-based multi-agent ecosystem (as of April 2026). The authors explicitly state these traces are observational and do not establish causal superiority over other regimes. The evidence illustrates five key patterns:

• Ungoverned Persistence Failure: An unmanaged auto-memory file contained three plausible false entries that persisted until a later audit, demonstrating how error becomes reloadable state without selection.
• Ratified Institutional Memory: A procedural lesson was successfully converted into a ratified principle, showing the outcome of the selection process.
• Version Registry Selection Loop: The system recorded a candidate version, identified defects, marked it failed, revised it, and then fixed the updated state. This demonstrates that rejection and revision are visible, not hidden.
• Identity-Preserving Expansion: A new agent joined the ecosystem, accessing shared institutional memory while maintaining separate private identity coordinates, validating the architectural separation of shared and local memory.
• Governance as Learning: A post-governance fabrication error occurred. Instead of preventing the error, the system documented the failure and converted it into a ratified "abstention-over-confabulation" principle, illustrating governance as a learning mechanism rather than a perfect firewall.

Key Contributions and Design Claims
The primary contribution is a design agenda for persistent LLM-based multi-agent systems. The paper argues that memory governance must be viewed as a selection layer that determines the system's future behavior. Key design claims include:

• Persistent systems require explicit selection regimes for shared memory.
• Governed collaborative memory must separate agent-local, shared institutional, archive, and project-continuity memories.
• Human-ratified artificial selection is essential when target properties include judgment quality, provenance fidelity, and role preservation.
• Rejected, abstained, and superseded memories must remain inspectable selection outcomes, not administrative noise.
• Evaluation of memory systems must expand beyond recall and performance to include provenance fidelity, selection traceability, epistemic quality, correction pathways, and role preservation.

Significance and Limitations
The paper's significance lies in reframing memory governance from a storage or access control problem to a selection regime problem. It posits that in persistent multi-agent systems, the decision of what becomes shared state is the most critical design choice, determining whether the system learns collectively or propagates errors.

The authors are modest in their claims, acknowledging that:

• The evidence is limited to one ecosystem and does not prove the superiority of human-ratified regimes over automatic ones.
• Human ratification introduces bottlenecks and potential operator bias.
• Scaling governance to multiple ratifiers or distributed authorities remains an unresolved challenge.
• Legacy memory systems face bootstrapping problems where pre-governance records may already influence behavior.

Ultimately, the paper calls for a shift in how designers and evaluators approach agent memory: moving from optimizing for retrieval efficiency to ensuring that the selection process itself is explicit, inspectable, and aligned with the preservation of heterogeneous agent identities.