D$^2$ACCI: New Dual-Loop Protocol for Diagnosing LLM Agent Memory Failures
A new paper on arXiv (2608.17756) introduces D$^2$ACCI, a diagnostic system that helps detect issues in persistent memory systems used by large language model agents. These memory systems allow agents to remember and adapt over multiple sessions, but their complex processes—like data ingestion and retrieval—make it hard to find where things go wrong. Traditional assessments only show that a problem happened, without identifying the specific stage involved. Current benchmarks lack detailed comparisons and diagnostics. D$^2$ACCI addresses this by using an outer gate to manage memory interventions based on evidence and monitoring. The paper also presents DCR, a metric for assessing failure localizability, and D$^2$ACCI-Eval, a tool designed for replaying these diagnostics, which is crucial for researchers studying LLM memory and evaluation.
Key facts
- D$^2$ACCI stands for Diagnostic-Driven Artifact-based Closed-loop Controlled Iteration
- D$^2$ACCI is a dual-loop protocol for localizing failures in LLM agent memory
- Persistent memory in LLM agents enables recall, revision, and personalization across sessions
- The memory pipeline consists of ingestion, retrieval, filtering, and generation stages
- End-to-end evaluation does not reveal which stage caused an error
- DCR is a graded observability metric measuring whether failures remain localizable
- D$^2$ACCI-Eval is a reusable artifact for gate replay
- The paper is on arXiv with ID 2608.17756, announced as new
Entities
—