Memory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents
DerivAudit checks whether long-term agent memories are supported by the history available when written.
The paper argues that persistent memories in long-running LLM agents can look unsupported when citations omit scattered evidence, or can overclaim by composing supported facts into stronger statements. DerivAudit checks whether the history available at write time actually supports what enters memory, separating missing citations, unsupported composition, and admission decisions. Across two natural memory corpora, broader pre-write history recovers support for nearly 60% of memories that appear unsupported from citations alone, while 17-21% remain unsupported. Expanding evidence does not make admission reliable, and on two backbones it worsens admission of unsupported memories.
- Audits evidence scope, compositional validity, and admission reliability
- Broader history supports nearly 60% of citation-unsupported memories
- 17-21% remain unsupported after evidence expansion
- Unsupported memories are still frequently admitted by verifiers
From the original article
1 min read at huggingface.co
Long-running LLM agents compress past interactions into persistent memories that may be reused as premises for later tasks. This creates a distinct derivation problem: whether the memory actually follows from what the interaction history supports. Relevant evidence may be scattered across earlier interactions, while compression can introduce relations or event status that the history never established.
By Hongjun Liu, Chen Zhao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.