ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-14

AI Researcher Florian_Dietz Proposes Agent Memory Audit Mechanisms: Preventing Goal Drift via Immutable Logs, Epistemic Status Markers, and User Sign-off

LessWrong author Florian_Dietz introduces a memory management approach designed to enhance the auditability of AI agents. The method relies on file conventions rather than complex software, implementing three core mechanisms: using topic-specific immutable LOG files as primary sources to prevent memory drift; introducing epistemic status markers like [decided] and [assumed] to distinguish hypotheses from confirmed facts; and requiring explicit user sign-off for important decisions, prohibiting silent agent approval. This proposal targets common risks in long-running agents, such as goal loss and unauthorized decision-making, offering engineering practices directly transferable to existing Obsidian or Wiki systems.

ImportanceMaterialEvidenceE3 inspectableWrite-upQuick

2026-10-14 21:42 GMT+8

When building AI agents with long-term memory, developers face a subtle but dangerous problem: as agents continuously edit and summarize their knowledge base, original information undergoes small, cumulative "drift," causing the agent to gradually deviate from its initial goals or make decisions based on hallucinations. AI researcher Florian_Dietz proposes EpistemicMemory, a file convention system published on LessWrong, which aims to solve this auditability issue through structured recording methods rather than relying solely on improved model capabilities.

Dietz points out that many existing memory tools (such as Karpathy's LLM Wiki or systems built on Obsidian) technically preserve original versions via Git history, but this is not agent-friendly because reviewing Git diffs requires significant computational effort and is often skipped. His solution involves creating per-topic chronological LOG files that are never altered. These files serve as immutable Primary Sources, detailing every conversation between the user and the agent. When the agent generates summaries or updates Wiki entries, it can always compare them against these raw logs, preventing semantic drift during multiple rewrites.

Sign in / Sign up
ReadingResearchRadarInvestment framework

To address the issue of agents confusing "plausibility" with "established fact," the system introduces Epistemic Status Markers. Agents must use specific tags when writing to memory, such as [decided], [verified], [concluded], [assumed], and [superseded]. This marking mechanism allows agents to compactly store the current status, alternative options, and the discussions leading to conclusions. For instance, when an old decision is overturned by new evidence, marking it as [superseded] and linking to the new discussion log clearly displays the reasoning chain, preventing the agent from mistakenly executing unverified assumptions as established facts.

The most critical anti-runaway mechanism is Explicit Signoff by the User. Dietz emphasizes that only human users have the authority to mark an idea as [decided]; the agent itself must never silently approve any important decision. If an agent proposes a suggestion, it must explicitly ask the user and receive a positive response before updating the status. This rule directly targets a common failure mode in long-running agents: the agent assumes silence equals consent and bypasses human supervision to proceed with tasks. By strictly limiting "approval rights" to human interaction, the system ensures that the agent's actions remain under human awareness and control.

This method is currently released as a GitHub template repository, consisting essentially of file naming conventions and prompt instructions, requiring no additional software installation. For developers already using other memory management systems, Dietz suggests having the agent scan the project and copy missing rules. Although the method depends on the agent's adherence to instructions and may fail under extreme context pressure, it offers a low-cost, high-transparency engineering path that allows humans to maintain audit capabilities over AI agents' internal states amidst trends toward automation.

Sources

  • 原始出处lesswrong.com↗
Continue researching →