When Not to Write Memory: Governing False Promotion from Correlated Agent Traces

Technical Summary: When Not to Write Memory: Governing False Promotion from Correlated Agent Traces

Problem Statement
As long-lived language agents increasingly generate reusable memories from their own execution traces, a critical safety failure mode has emerged: false promotion. The core issue is that repeated observations across multiple agents are not necessarily independent evidence. A claim repeated by five agents may represent five independent discoveries, or it may represent a single stale note, a shared prompt artifact, a copied source, or a tool-induced echo. Naively promoting such correlated repetitions into long-term memory creates a persistence layer for correlated errors. The paper frames the challenge not as what to remember, but when to refuse writing memory, specifically addressing the governance of write paths for correlated multi-agent traces.

Methodology: GovMem
The authors introduce GovMem, a conservative diagnostic reference policy and side-agent auditor designed to govern memory write paths. Rather than treating memory writing as a summarization problem, GovMem frames it as an evidence-governance decision. The system operates as a task-independent auditor that intercepts candidate memories before they reach long-term storage.

The GovMem pipeline executes four steps:

1. Candidate Aggregation: It aggregates observations with similar claim text while retaining provenance metadata (source, prompt family, parent event, environment, trust).
2. Dependency-Aware Support Estimation: Instead of counting raw frequencies or unique sources, it groups observations by dependency structure. Observations sharing prompts, tools, intermediate context, or parent events are treated as non-independent and do not count as separate votes.
3. Counterexample Retrieval & Scope Assignment: It retrieves potential counterevidence and checks target-scope compatibility, recognizing that textual agreement alone is an uncalibrated confidence signal.
4. Decision Routing: Based on dependency-aware verification, it outputs one of three decisions: PROMOTE, REJECT, or NEEDS-REVIEW.

Key Contributions

1. Failure Mode Identification: The paper identifies "false promotion" from copied, prompt-correlated, stale, adversarial, and out-of-scope observations as a distinct and critical memory failure mode.
2. Governance Formulation: It reframes memory writing as an evidence-governance problem, routing candidates to promote, reject, or review based on provenance and dependency rather than summarization quality.
3. GovMem Implementation: It instantiates this as a conservative side-agent policy that audits write proposals using dependency-aware support, counterevidence, and scope checks.
4. Comprehensive Evaluation: The authors provide controlled mechanism experiments (GovMem-Bench), a project-internal human-gold real-trace evaluation, and external stress tests, including a final human-adjudicated high-impact packet.

Experimental Results
The evaluation spans synthetic benchmarks, internal real traces, and external coding-agent stress tests.

• Controlled Synthetic Tests (GovMem-Bench): In default settings, GovMem-Verified reduced the false promotion rate from 0.597 (raw frequency baseline) to 0.040, while maintaining a recall of 0.960. This came at an explicit review burden of 0.151.
• Internal Real-Trace Evaluation: On a human-labeled subset of 120 candidates from 79 traces, dependency-aware promotion reduced false promotion from 0.371 (source+scope baseline) to 0.032. However, this safety gain required a significant increase in review burden (from 0.042 to 0.692) and a drop in direct recall (from 0.985 to 0.448).
• External Stress Tests (V2 High-Impact Packet): A final human adjudication of 133 high-impact external coding-agent candidates yielded a severe result: zero candidates were deemed safe for automatic promotion. Notably, all 11 candidates that passed a local verification gate were rejected by human adjudicators as boilerplate, shared-tool artifacts, file dumps, or non-reusable debugging traces.
• Baseline Comparisons: A "review-all-correlated" baseline (routing all non-independent evidence to humans) performed surprisingly well, nearly matching GovMem's safety with better recall and lower burden in the internal setting, suggesting GovMem is a specific design point in a risk-control space rather than a universally dominant solution.

Significance and Claims
The paper explicitly positions itself as a diagnostic and method paper rather than a validation of a generally efficient automatic memory writer.

• Diagnostic Safety Claim: The primary contribution is demonstrating that governed write paths, which treat repetition as correlated evidence requiring provenance and scope checks, can significantly reduce false promotion.
• Limitations on Automation: The authors caution that automatic promotion remains brittle when verification evidence is extracted from noisy agent traces. The external results show that surface verification cues (e.g., pytest output, file listings) are often insufficient for safe memory writes.
• Design Space: GovMem is presented as one point in a risk-control design space, emphasizing explicit provenance and conservative decisions. The paper argues that agent memory write paths should be evaluated as risk-controlled evidence-governance systems.
• Future Requirements: The authors state that stronger claims regarding automatic memory writing require broader independently labeled external coverage, executed downstream-harm evidence, and stronger event-level verification, none of which are fully established in this work.

In summary, the paper argues that "repetition is not evidence" for long-term memory and proposes a governance framework to filter correlated errors, while acknowledging that current automatic promotion mechanisms are not yet robust enough for deployment without significant human review or further validation.