Reproduce “Always-OnAgents:A Survey of Persistent Memory, State, and Governance in LLMAgents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianyu Ding, Aditya Nannapaneni, Bingfan Liu, Ling Zhang
This paper defines always-on agents as persistent-state LLM systems, proposing a framework with a 10-stage lifecycle and five invariants for robust state governance. It introduces the Always-On Evaluation Protocol (AOEP-v0) which reveals that existing memory systems often fail critical governance obligations despite strong recall.
Key insights
- Current research exhibits a significant imbalance, heavily emphasizing state accumulation and retrieval ("forward arc") while neglecting governance, recovery, and relinquishment ("return arc") operations.
- Existing memory augmentation mechanisms primarily optimize for accumulating competence (plasticity) but often lack robust stability operators like validation, audit, forgetting, and rollback.
- Dominant memory substrates (e.g., vector stores) are efficient for write/read but inherently lack native support for explicit governance features like authority, recoverability, and comprehensive deletion propagation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.