RESEARCH · RESEARCH · #1240
arXiv paper 'Memory Is a Derivation' (v1) introduces DerivAudit to audit LLM agent memories
The new arXiv preprint 'Memory Is a Derivation' (arXiv:2609.36130v1) identifies a derivation problem for long-running LLM agents whose persistent memories compress past interactions, formalizing three requirements (evidence scope, compositional validity, admission reliability) and introducing DerivAudit to test whether a stored memory is actually supported by the history available at write time. Audits on two memory corpora show that expanding pre-write history recovers support for nearly 60% of memories that appear unsupported from citations alone, 17–21% remain unsupported after expansion, and broader evidence can worsen admission reliability on some model backbones.
KEY POINTS
- The new arXiv preprint 'Memory Is a Derivation' (arXiv:2609.36130v1) identifies a derivation problem for long-running LLM agents whose persistent memories compress past interactions, formalizing three requirements (evidence scope, compositional validity, admission reliability) and introducing DerivAudit to test whether a stored memory is actually supported by the history available at write time.
- Audits on two memory corpora show that expanding pre-write history recovers support for nearly 60% of memories that appear unsupported from citations alone, 17–21% remain unsupported after expansion, and broader evidence can worsen admission reliability on some model backbones.
- This matters because it highlights a concrete risk to the correctness and auditability of long-term LLM agent memory—affecting downstream behavior, trust, and verification of agents.
WHY IT MATTERS
This matters because it highlights a concrete risk to the correctness and auditability of long-term LLM agent memory—affecting downstream behavior, trust, and verification of agents.