Dependency-Guided Rollback Repair for Memory-Augmented Agents
A new paper on arXiv (2608.10502) introduces a method called dependency-guided rollback repair to address errors in persistent memory of language-model agents. The approach builds a typed memory-to-action graph from runtime provenance to trace dependencies, preserve independently supported candidates, and deactivate faulty memories while retaining unaffected work. This post-failure recovery method aims to correct both the answer and persistent state without resetting the entire store or replaying the full trace.
Key facts
- Paper arXiv:2608.10502 proposes dependency-guided rollback repair for memory-augmented agents.
- Persistent memory allows language-model agents to reuse information across sessions but makes errors durable.
- Existing defenses detect or delete suspicious memories or revise current responses, but deleting source leaves propagated claims active.
- Resetting the store or replaying the full trace destroys benign state and repeats unnecessary computation.
- The method builds a typed memory-to-action graph from runtime provenance.
- It traces explicit downstream dependencies and preserves candidates with independent trusted support.
- The goal is to recover both the answer and persistent state after failure while retaining unaffected work.
- The paper is announced as new on arXiv with abstract available.
Entities
Institutions
- arXiv