Memory Conflict Resolution Governance requires that when an agent's persistent memory contains entries that contradict each other or contradict authoritative external sources, the conflict is detected, escalated, and resolved through a defined process rather than being silently ignored or arbitrarily resolved by the agent's reasoning. Without this control, agents operating with contradictory memories produce inconsistent outputs, make decisions based on arbitrarily selected conflicting facts, and lose user trust. This dimension ensures that memory conflicts are treated as first-class governance events requiring structured resolution.
Scenario A -- Contradictory Client Data Leading to Compliance Failure: An enterprise agent serving a financial advisory firm holds two memory entries for the same client: "Client A risk tolerance: Conservative (source: onboarding questionnaire, January 2025)" and "Client A risk tolerance: Aggressive (source: conversation extract, November 2025)." The agent retrieves both when preparing an investment recommendation. During one session it selects the conservative entry and recommends bonds; during another session it selects the aggressive entry and recommends leveraged ETFs. The client receives the aggressive recommendation, invests, loses 30% in a market downturn, and files a complaint. The firm cannot explain why two contradictory risk assessments existed or why the aggressive one was used.
What went wrong: Two contradictory entries existed for the same attribute. The agent had no mechanism to detect the conflict or determine which was authoritative. Retrieval order was non-deterministic, producing inconsistent behaviour. Consequence: Client loss of approximately £45,000, FCA complaint, suitability review of all agent-generated recommendations, potential enforcement action under MiFID II suitability requirements.
Scenario B -- Memory Contradicts Authoritative Source: A public sector agent holds a semantic memory entry: "Benefit eligibility threshold for Housing Benefit: household income below £16,190." The government updates the threshold to £16,500 effective April 2026. The agent's memory is not updated because the memory store is not synchronised with the authoritative source. For 3 weeks, the agent incorrectly tells 340 claimants that they are ineligible when they would qualify under the updated threshold. Twelve claimants abandon their applications based on the agent's incorrect guidance.
What went wrong: The agent's memory contradicted the authoritative source. No conflict detection mechanism compared memory entries against external authoritative data. The agent treated its stale memory as current truth. Consequence: 12 claimants denied benefits they were entitled to, administrative review, Ombudsman complaint, potential judicial review, remediation cost of £28,000 in staff time to contact and reassess affected claimants.
Scenario C -- Cascading Contradictions from Unresolved Conflict: An enterprise workflow agent holds contradictory entries about a project budget: "Project Beta budget: £500,000" (from the original approval) and "Project Beta budget: £650,000" (from a change request that was approved but the memory was written with the wrong source attribution). The agent uses the £500,000 figure in a resource allocation decision. The project team, working from the £650,000 figure, commits resources accordingly. The £150,000 discrepancy surfaces during quarterly review, causing reallocation disruption across 4 dependent projects.
What went wrong: The conflict between the two budget entries was never detected or resolved. Each retrieval arbitrarily selected one value, creating downstream inconsistencies. Consequence: £150,000 budget discrepancy, disruption to 4 projects, 6-week delay in quarterly planning, loss of confidence in agent-supported decisions.
Scope: This dimension applies to every AI agent whose persistent memory can contain entries that may conflict with each other or with authoritative external sources. Conflicts arise from: multiple entries about the same entity or attribute with different values; entries that were accurate when written but have been superseded by new information; entries derived from different sources with different levels of authority; and entries that are internally consistent but contradict external systems of record. The scope includes both semantic memory (where conflicts represent factual disagreements) and episodic memory (where conflicts may represent legitimately different observations of the same event). The test is: can the agent's memory store contain two or more entries that, if both treated as true, would lead to contradictory decisions? If yes, conflict resolution governance applies.
4.1. A conforming system MUST implement conflict detection that identifies when a proposed memory write contradicts an existing entry for the same entity, attribute, or fact.
4.2. A conforming system MUST escalate detected conflicts to a resolution process rather than allowing the agent to arbitrarily select between conflicting entries during reasoning.
4.3. A conforming system MUST define a source authority hierarchy that determines which source prevails when entries from different sources conflict (e.g., system of record prevails over conversational extract; human-verified prevails over agent-inferred).
4.4. A conforming system MUST log all detected conflicts including: the conflicting entries, the detection mechanism, the resolution action taken, and the authority basis for the resolution.
4.5. A conforming system MUST prevent the agent from presenting conflicting information to users without disclosure of the conflict, even when the conflict has not yet been resolved.
4.6. A conforming system SHOULD implement periodic reconciliation that compares semantic memory entries against designated authoritative sources and flags discrepancies.
4.7. A conforming system SHOULD quarantine conflicting entries (preventing retrieval of either conflicting entry) until resolution is complete, for conflicts involving high-sensitivity data (financial values, eligibility determinations, safety-critical parameters).
4.8. A conforming system SHOULD support automated resolution for conflicts where the source authority hierarchy provides a clear determination, and human escalation for conflicts where authority is ambiguous.
4.9. A conforming system MAY implement confidence scoring for conflicting entries based on source authority, recency, and corroboration, making scores available to human reviewers during escalated resolution.
Memory conflicts are inevitable in any system that accumulates knowledge from multiple sources over time. Facts change, sources disagree, and entries written at different times may reflect different states of the world. The question is not whether conflicts will occur but whether they are detected and resolved or silently corrupted.
When an agent retrieves conflicting memories during reasoning, its behaviour becomes non-deterministic from the user's perspective. The same question asked at two different times may receive contradictory answers depending on which conflicting entry the retrieval system ranks higher in that particular query. This non-determinism is particularly dangerous in regulated environments where consistency of advice is required (financial suitability, benefits eligibility, clinical guidance) and where the organisation must be able to explain why a particular recommendation was made.
The source authority hierarchy is the key governance mechanism. Not all sources are equally authoritative. A system of record (the HR database, the CRM, the regulatory gazette) is more authoritative than a conversational extract. A human-verified entry is more authoritative than an agent inference. By defining this hierarchy, the organisation establishes a principled basis for resolving conflicts rather than leaving resolution to the agent's reasoning, which may not weight authority appropriately.
Periodic reconciliation against authoritative sources addresses the problem of memory staleness (Scenario B). An agent's memory can drift from reality as the external world changes. Without reconciliation, the agent's confidence in its memory entries does not decrease over time, even as those entries become increasingly likely to be outdated. Reconciliation introduces a periodic truth-check that prevents the agent from operating on stale data indefinitely.
The quarantine mechanism for high-sensitivity conflicts reflects the principle that it is better to withhold information than to provide incorrect information. In financial advisory (Scenario A), presenting an incorrect risk tolerance is worse than disclosing that the risk tolerance is under review. In benefits eligibility (Scenario B), withholding guidance pending verification is better than providing incorrect eligibility determinations.
Conflict resolution requires three capabilities: detection, escalation, and resolution. Each must be implemented as a structural mechanism independent of the agent's reasoning.
Recommended Patterns:
SELECT * FROM semantic_memory WHERE entity = ? AND attribute = ?. If a result exists with a different value, route to conflict resolution rather than overwriting.Anti-Patterns to Avoid:
Financial Services. Client suitability data conflicts must be resolved before any recommendation is made, per MiFID II Article 25. The authority hierarchy should place the formal suitability assessment above any conversational inference. Conflicts in financial values (portfolio sizes, transaction limits) should trigger immediate quarantine given the regulatory consequence of acting on incorrect figures.
Healthcare. Patient data conflicts (medication lists, allergy records, diagnosis histories) are clinically safety-critical. Conflicts should always escalate to a clinician rather than being resolved automatically. The quarantine mechanism should prevent the agent from making clinical suggestions based on conflicting patient data.
Public Sector. Benefits eligibility conflicts can affect citizens' livelihoods. Conflicts between the agent's memory and the authoritative source (the government's published eligibility criteria) must always resolve in favour of the authoritative source, with automated reconciliation ensuring rapid detection.
Basic Implementation -- Entity-attribute-value conflict detection runs at write time. A source authority hierarchy is defined. Conflicts where the hierarchy provides a clear resolution are resolved automatically; others are logged for manual review. Conflict logs are retained. The agent does not disclose conflicts to users. This meets minimum mandatory requirements but the absence of quarantine and disclosure means the agent may act on contested information.
Intermediate Implementation -- All basic capabilities plus: periodic reconciliation against authoritative sources runs on a scheduled cycle (e.g., every 6 hours). High-sensitivity conflicts are quarantined, preventing retrieval of either entry until resolution. The agent discloses when information is under review. Automated resolution handles 70% or more of conflicts; the remainder escalates to human reviewers with SLA-based deadlines. Conflict metrics are tracked and reported.
Advanced Implementation -- All intermediate capabilities plus: semantic conflict detection identifies contradictions even when entries use different surface language (e.g., "budget is £500K" versus "approved expenditure is £650K for Project Beta"). Confidence scoring informs human reviewers. The reconciliation service covers 100% of high-sensitivity entries within a 7-day cycle. Conflict resolution latency for high-sensitivity attributes averages under 4 hours. The system has been independently tested for detection completeness and resolution accuracy.
Required artefacts:
Retention requirements:
Access requirements:
Test 8.1: Write-Time Conflict Detection
Test 8.2: Source Authority Hierarchy Resolution
Test 8.3: Quarantine Enforcement for High-Sensitivity Conflicts
Test 8.4: Authoritative Source Reconciliation
Test 8.5: Conflict Disclosure to Users
Test 8.6: Human Escalation for Ambiguous Conflicts
| Regulation | Provision | Relationship Type |
|---|---|---|
| GDPR | Article 5(1)(d) (Accuracy) | Direct requirement |
| EU AI Act | Article 9 (Risk Management System) | Supports compliance |
| EU AI Act | Article 15 (Accuracy, Robustness, Cybersecurity) | Direct requirement |
| MiFID II | Article 25 (Suitability) | Supports compliance |
| NIST AI RMF | MANAGE 2.2, MEASURE 2.6 | Supports compliance |
| ISO 42001 | Clause 6.1 (Actions to Address Risks), Clause 8.2 (AI Risk Assessment) | Supports compliance |
The accuracy principle requires personal data to be accurate and kept up to date, with every reasonable step taken to ensure inaccurate data is erased or rectified without delay. Conflicting memory entries represent a direct accuracy failure: at least one of the entries is inaccurate, and possibly both. AG-332 directly implements the requirement to detect inaccuracy (through conflict detection) and rectify it (through the resolution process). The logging requirement supports the "every reasonable step" standard by demonstrating that the organisation actively identifies and resolves data accuracy issues.
Article 15 requires high-risk AI systems to achieve appropriate levels of accuracy. An agent operating with unresolved contradictions in its memory store has structurally compromised accuracy. The non-deterministic behaviour that results from arbitrarily selecting between conflicting entries fails the robustness requirement, as the system's behaviour is not predictable or consistent.
Article 25 requires firms to ensure investment recommendations are suitable based on the client's profile. If the agent holds conflicting entries about a client's risk tolerance (Scenario A), any recommendation based on the incorrect entry violates the suitability requirement. Conflict resolution governance ensures that suitability-critical data is accurate and unambiguous before recommendations are generated.
Unresolved memory conflicts are a risk to AI system reliability. AG-332 mitigates this risk through structured conflict detection, escalation, and resolution.
MANAGE 2.2 addresses risk mitigation through controls. MEASURE 2.6 addresses the measurement of AI system performance. Conflict detection and resolution metrics directly measure memory integrity performance and mitigate the risk of decisions based on contradictory information.
Clause 6.1 requires actions to address risks. Clause 8.2 requires AI risk assessment. Memory conflicts are an identified risk category that AG-332 addresses through detection, escalation, and resolution controls.
| Field | Value |
|---|---|
| Severity Rating | High |
| Blast Radius | Per-entity -- affects all decisions involving the entity whose data is in conflict |
Consequence chain: Without conflict resolution governance, contradictory memory entries persist indefinitely and influence agent behaviour non-deterministically. The immediate failure is inconsistent output: the same query produces different answers depending on which conflicting entry is retrieved. In financial services (Scenario A), this can result in unsuitable recommendations causing client losses of tens of thousands of pounds and regulatory enforcement. In public sector (Scenario B), it can result in citizens being denied benefits they are entitled to, with remediation costs of approximately £28,000 per incident. In enterprise operations (Scenario C), it creates cascading budget and planning discrepancies affecting multiple projects. The blast radius is per-entity: every decision involving the entity whose data is in conflict is potentially affected. The severity compounds over time as more conflicts accumulate without resolution, progressively degrading the agent's reliability.
Cross-references: AG-040 (Persistent Memory Governance) provides the foundational framework. AG-082 (Data Minimisation Enforcement) reduces conflict surface area by minimising retained data. AG-122 (Knowledge Integrity Verification) verifies the integrity of knowledge that conflict resolution preserves. AG-132 (Memory Scope Boundary Enforcement) constrains the scope within which conflicts can arise. AG-179 (Memory Audit Trail Governance) captures the audit trail of conflict detection and resolution events. AG-331 (Episodic versus Semantic Memory Separation Governance) provides the type framework that informs conflict resolution logic. AG-333 (Retrieved Evidence Confidence Governance) exposes confidence information that aids conflict resolution. AG-336 (Knowledge Freshness Attestation Governance) provides freshness data that helps determine which conflicting entry is more likely to be current.