The author argues that AI agents have a state-integrity problem rather than a memory issue, proposing a State Ledger to distinguish historical facts from current state and track provenance.
I’ve been building a long-running, model-agnostic AI system, and I kept running into a failure that initially looked like a memory problem. It wasn’t. The system could correctly remember something that happened weeks ago. The problem was that the thing it remembered was no longer true. Here’s a simplified version of an actual failure mode I encountered: Earlier: "Use architecture X." Later: "Changed the architecture. Use Y instead." Much later: Agent retrieves: "Use architecture X." Nothing was necessarily wrong with retrieval. The old information was real. It was just historically true rather than currently true. That’s the distinction I’m now experimenting with: MEMORY = something that happened STATE = what is currently believed to be true PROVENANCE = why it is believed to be true So instead of treating memory as a collection of facts, I’m experimenting with representing change explicitly: OBSERVATION ↓ INTERPRETATION ↓ HYPOTHESIS ↓ EXPERIMENT ↓ OUTCOME ↓ STATE CHANGE And attaching provenance to claims: VERBATIM_AVAILABLE CONTEXT_AVAILABLE DERIVED UNKNOWN Along with relationships such as: supersedes contradicts derived_from validated_by So the old decision isn’t deleted. It’s preserved as history: Decision A ↓ superseded by ↓ Decision B ↓ current state The agent can then distinguish: “I remember that X was previously decided.” from: “X is the current decision.” The harder problem I’ve found is that inference itself can become memory. An agent sees several observations, infers a pattern, stores that pattern, and eventually retrieves the inference as though it were an original fact. That’s the failure I’m trying to prevent. The goal isn’t to make an agent remember everything. It’s to make it preserve what happened, what changed, what is currently believed, why it is believed, and what remains uncertain. I’m calling the experiment a State Ledger. For people building long-running agents: When stored memory conflicts with newer evidence, what mechanism does your system use to establish the current state and can you trace that state back to the evidence that established it?
The author explores the idea of AI agents having a public, auditable memory to record important decisions, which could enhance trust but also introduce new complexities.
The article argues that most production failures in AI agents are due to unstable operational state and memory degradation, not weak models, and emphasizes the need for better infrastructure for state management, observability, and adaptive reliability.
A reflection on agent memory as primarily an infrastructure/data-management problem rather than an AI problem, focusing on practical complexities like permissions, scopes, and revision history.
The article critiques current AI memory systems as mere write-only logs that lack the ability to be corrected, updated, or traced to their source, arguing that true memory requires a governance layer.
Reflects on the continuity problem for long-running AI agents, arguing that a deterministic control layer is needed to manage authoritative state, and questions whether existing infrastructure like IAM, transactions, and provenance is sufficient.