Memory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents
Hongjun Liu, Chen Zhao
New York University
Published September 28, 2026 · on the feed September 30, 2026 · ▲ 6 · arXiv 2609.36130
The abstract
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. A valid memory may therefore appear unsupported because its citations omit relevant evidence, while individually supported facts may be composed into a stronger statement the history never established. We characterize this problem through three coupled requirements: (1) Evidence scope; (2) Compositional validity; (3) Admission reliability. We therefore ask whether the interaction history available at write time supports what enters persistent memory. We introduce DerivAudit, a framework for auditing whether a memory is actually supported by the history available when it was written. The audit separates three questions: whether supporting evidence lies beyond writer-provided citations, whether the composed memory introduces unsupported meaning, and how write-time admission decisions affect later memory use. Across two natural memory corpora, audits using broader pre-write history recover support for nearly 60% of memories that appear unsupported from citations alone, while 17-21% remain unsupported after expansion. Yet broader evidence does not by itself make admission reliable: unsupported memories are still frequently admitted across verification models, and evidence expansion alone worsens it on two backbones.
The Brief
2:29A short podcast summary of the paper. Written and voiced by AI.
Transcript
Memory Is a Derivation. The Distributed-Evidence Paradox in Long-Term Agents. By Hongjun Liu and one other, from New York University. Published on the twenty-eighth of September. No code repository is linked to it.
Long-running LLM agents compress their past interactions into persistent memories, and those memories may later be reused as premises for new tasks. That creates a problem of derivation: does a memory actually follow from what the interaction history supports? Two things make the question hard. The relevant evidence may be scattered across many earlier interactions. And compression can introduce relations, or claims about whether an event happened, that the history never established.
The difficulty cuts both ways, producing a pair of opposite errors. A valid memory can look unsupported, because its citations leave out evidence that does exist. Meanwhile facts that are each supported can be combined into a stronger statement that the history never established. The authors frame the problem as three coupled requirements: evidence scope, compositional validity, and admission reliability. The question they ask is whether the history available at the moment of writing supports what enters persistent memory.
DerivAudit is a framework for auditing exactly that, whether a memory is supported by the history available when it was written. It splits the audit into three questions, one per requirement. Does supporting evidence lie beyond the citations the writer provided? Does the composed memory introduce meaning that nothing supports? And how do decisions made at write time, about what to admit, affect how memory is used later?
The audits run on two natural memory corpora. A key contrast is between judging a memory from its citations alone and judging it against the broader history that existed before it was written. Admission is then examined across several verification models and backbones.
Widening the evidence changes the picture a lot. Audits that draw on the broader pre-write history recover support for nearly sixty percent of the memories that looked unsupported from their citations alone. Even after that expansion, seventeen to twenty-one percent remain unsupported.
More evidence does not by itself fix admission. Unsupported memories are still admitted frequently, across verification models. And on two backbones, expanding the evidence alone makes admission worse.
Citations turn out to be a poor proxy in both directions. They undersell memories that are sound and fail to catch ones that overreach. Treating each memory as a derivation to be checked against its full history gets closer to the truth, but deciding what to let in remains its own unsolved step.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
An agent's memories become its facts. What if some were never true?
You build a long-running agent that compresses past conversations into stored memories. Later tasks treat those memories as facts and never look back at the original history.
The road ahead
- 1 · The problem ~7 min An agent's stored memories become facts for later tasks, but checking them against cited passages both rejects valid memories and misses stitched-together false claims.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Treat each memory as a conclusion that must follow, as a whole, from the history available when it was written.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Gather real memory writes, judge each from citations alone, rejudge it with wider earlier history, then test which memories different verifiers admit.
- 6 · The experiment ~6 min You’ll guess: Of the memories that looked unsupported from their citations alone, roughly what share had support recovered once the audit used the broader earlier history?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.