When several agents remember different things, who wins?
Give several agents one shared memory and they will eventually collide: two writers, one key, incompatible claims. This is the follow-up promised in Provenance in agent memory — provenance tells a system who asserted a claim; it does not decide which claimant is right.
TL;DR
- Majority voting is vulnerable to cheap identities and copied claims. Current RAG systems can still follow the majority when a more credible source disagrees.
- In the surveyed systems, reputation did not remove authority — it moved it into the start vector (EigenTrust) or into accumulated stake (UMA).
- Four bodies of published work resolve some disputes without ranking people: outcome scoring, meta-knowledge, independence detection, and reason-based aggregation.
- Conflict detection can stay coordination-free. Final resolution sometimes cannot — so keeping both versions may be the honest design choice.
A multi-agent memory conflict happens when several writers put incompatible claims into one shared, persistent memory.
The tempting rule is simple: count the agents and keep the majority answer. That rule fails when identities are cheap, evidence is correlated, or agents repeat what they read.
Why majority voting can fail in multi-agent memory
Douceur’s Sybil attack analysis showed that identities are nearly free to create unless someone certifies them centrally. A rule that counts writers is captured by minting more writers.
This is not only a theoretical concern. CONFACT, published at IJCAI 2025, found that state-of-the-art RAG systems tend to follow the majority of retrieved contexts on conflicting-evidence questions. They do so even when one stronger source contradicts several weak ones.
Standard truth-discovery methods hide the same default. Start every source at equal weight, and the algorithm's first step is majority voting — the survey literature says so directly (survey). Reliability scores come later, learned from agreement; the majority supplied the starting point.
Debate does not reliably repair the problem. In “Talk Isn’t Always Cheap”, models frequently changed correct answers to incorrect ones after seeing peer reasoning, favoring agreement over challenging an error. A separate evaluation, “Should we be going MAD?”, found that current multi-agent debate does not reliably outperform self-consistency or ensembling. One emerging response is to preserve disagreement inside the architecture rather than force consensus — argued so far mainly for value-based deliberation, where no objectively correct answer exists (preprint).
Consensus can be a symptom of social pressure rather than stronger evidence.
Why reputation systems can recreate authority
A reputation system weights a source’s future claims by a score built from its past ones.
Reputation seems to avoid central judgment: the system computes the ranking. But the computation still needs a place to start.
The EigenTrust paper describes its set of pre-trusted peers as “essential.” Those peers are the authority, sitting in the start vector. Cheng and Friedman went further: in their model, anonymous Sybil-proof reputation simply does not exist (P2PECON 2005).
Transferable influence creates another failure mode. On March 24–25, 2025, one large UMA holder cast about 5 million UMA across three accounts — roughly 25% of the dispute-round votes. A ~$7 million Polymarket market on a Ukraine mineral deal then resolved “Yes” against reality; Polymarket called it an “unprecedented” governance attack (reported consistently across crypto press — The Block, CoinMarketCap Academy).
The subsequent fix made the authority explicit. UMIP-189, proposed July 29, 2025 and adopted that August, restricted who may submit a resolution proposal — a proposer whitelist, reported by crypto press as starting at 37 vetted addresses and reaching 177 by November 2025, while the right to dispute stayed open to anyone.
The incident does not establish that every stake-based system will fail. It demonstrates the plutocratic failure mode: accumulated influence can dominate a factual dispute, after which the repair reinstalls a recognized set of authorities. In this incident, durable transferable stake did not eliminate authority — it changed the name on the door.
Agreement without independence adds little evidence
Goldman’s analysis of expert testimony makes a simple point: several agreeing sources add no more weight than one, unless they are at least partly independent of each other (“Experts: Which Ones Should You Trust?”). A source that merely echoes another adds no confirmation.
Experimental evidence points in the same direction. In a study of 144 subjects across 12 sessions, Lorenz and colleagues found that mild social influence narrowed the range of answers and hurt the crowd’s accuracy — while confidence went up (PNAS, 2011).
Agent pools are correlated by construction when they share a base model, prompts, retrieved contexts, or memory reads. An agent that reads a claim from shared memory and asserts it again is a copier, not an independent witness. When every source is a child of the same base model, the crowd is a chorus of one.
Copy detection offers a concrete answer. Dong, Berti-Équille, and Srivastava catch dependence through shared false values: shared truths reveal little, but shared errors are a fingerprint of copying. Their method cuts a copier’s vote toward zero and keeps the full weight of a small independent source (PVLDB 2009).
The relevant unit is not the agent identifier but the independently produced evidence behind the claim. Independence, not authority, is the currency.
Four published approaches to authority-free conflict resolution
Authority-free conflict resolution settles disputes without ranking the people or agents who make the claims.
Four bodies of published work illustrate different ways to do this. They are alternatives for different conditions, not a claim that every dispute fits one of them.
| Published family | What the mechanism uses | Representative result |
|---|---|---|
| Outcome anchoring | Eventually verifiable outcomes | Proper scoring rules reward calibrated reports (Brier, 1950; Gneiting and Raftery, 2007). Prediction-with-expert-advice methods update weights from observed results rather than prior source rank. |
| Meta-knowledge | Beliefs about other respondents’ beliefs | Surprisingly Popular chooses an answer that is more popular than respondents predicted. |
| Independence evidence | Evidence that reports were produced separately | Copy detection discounts derivative reports; Dawid–Skene models estimate observer error rates from repeated judgments. |
| Structure of reasons | Premises, attacks, and justification graphs | Premise-based judgment aggregation evaluates supporting propositions. In the studied argumentation setting, nomination lets one dissenter register a conflict (Chen & Endriss, TARK 2017). |
The most counterintuitive result comes from meta-knowledge.
The Surprisingly Popular method asks for both an answer and a prediction of how others will answer. It selects the answer that turns out more popular than predicted. An informed minority can win because its members know that their answer will be underestimated.
In the reported experiments, the method reduced errors by 21.3% relative to majority voting and by 24.2% relative to confidence-weighted voting (MIT summary); on the 50-state-capitals set it reduced incorrect decisions by 48% (Princeton Engineering; Nature 541:532). The canonical example contrasts the widely guessed Philadelphia with the correct capital, Harrisburg.
The mechanism uses no identity information. Meta-knowledge, rather than writer rank, separates the informed minority from the confident majority.
A related mechanism, Bayesian Truth Serum, addresses subjective questions where objective truth is unknowable. It rewards honest reporting, including reports from respondents who believe they hold a minority view. It protects candor; it does not select an objective truth.
Conflict detection is easier than final resolution
The CALM theorem states that a program has a consistent, coordination-free distributed implementation if and only if it is monotonic.
A growing set of discovered contradiction pairs is monotonic. New evidence can add conflicts without invalidating earlier detections. Declaring a final winner is different: it may require knowing that no defeating claim exists anywhere else.
Production systems already live by this split. Amazon’s Dynamo returns conflicting versions as siblings and leaves the retained-version decision to the application. Concurrent writes create the divergence; the system surfaces it to the client at read time and leaves resolution to the reader’s context.
For shared agent memory, that suggests a practical stance:
- Detect incompatible claims whenever they appear.
- Resolve them when an external outcome can referee.
- Use meta-knowledge when a factual answer exists but direct verification does not.
- Discount agreement when provenance indicates copying or shared error sources.
- Preserve sibling claims when the evidence does not justify a winner.
Keeping both sides is not indecision. Refusal — "we don't know yet" — is a design choice the literature supports, not an admission of failure; in some disputes it is the most truthful available output.
Mnemoverse and multi-writer memory
Rooms, currently in beta, provide a membership-checked memory pool for agents under different accounts. Reads carry authorship through provenance fields and author filters. That supplies the identity half of conflict analysis.
Mnemoverse’s feedback mechanism scores claims through task outcomes marked helpful or unhelpful (how the feedback loop works). It reads a claim’s own history, not the writer’s rank, and the engine caps how far feedback can move a ranking.
Adjudication between conflicting writers remains an open research problem. This article states the published basis for that work; it does not claim that Mnemoverse has solved it.
Common questions
Can AI agents resolve conflicts without a central authority?
Yes, in some settings. Published mechanisms can use verifiable outcomes, meta-knowledge, independence evidence, or the structure of reasons instead of ranking writers.
What is the Surprisingly Popular algorithm?
The Surprisingly Popular algorithm selects the answer that gets more support than people predicted. An informed minority can beat the majority — with no identities and no reputation scores.
Why can reputation systems get captured?
In the systems surveyed here, reputation needed a starting trust rule, and influence that could be stored and moved piled up. EigenTrust needed pre-trusted peers; in UMA voting, accumulated stake dominated a real dispute.
Why can majority voting fail for multi-agent memory?
Majority voting counts identities rather than independent evidence. Cheap Sybil identities, copied claims, and correlated agents can manufacture apparent consensus.
Why can agreement between AI agents be worthless?
Agreement adds little evidence when agents share error sources or copy one another. Distinct agent identifiers do not establish conditional independence.
Related
- World representation for AI agents
- Provenance in agent memory
- Bitemporal memory for AI agents
- Shared-memory poisoning
- Graph memory and MCP tools
- Rooms API
Edward Izgorodin · Mnemoverse · 2026-07-25
— Mnemoverse is a persistent-memory API for AI agents. Free key: console.mnemoverse.com · Docs: Getting Started