When Old Facts Return: Re-Reads, Reverts, and the Limits of Temporal Memory
- Source
- arxiv.org
- Author
- Neeraj Yadav (Called It Inc.)
- Date

Context
A coding assistant's memory should know when a fact has been replaced, such as a renamed . The paper, submitted October 6, 2026, points to an ambiguity: a source being read again and a genuine revert can produce the same sequence of observed values, yet require opposite current answers.
On 130 extractor-selected transitions derived from software fixes, an identity-based temporal memory reached 98.5 percent -judged accuracy in the ordinary condition. After appending a verbatim re-read of the old statement, accuracy fell to 10.8 percent and the stale-value rate rose to 88.5 percent. A guard that refuses to reactivate a retired value restored accuracy to 97.7 percent in this constructed condition, but it cannot recognize a legitimate revert without additional change provenance.
The author limits conclusions to the retained experiments, citing selected inputs and related-family judges. An earlier study by the same author also used 130 atomic transitions.
- The core ambiguity: re-reading an old source and a genuine revert can produce the same observed sequence of values, yet they require opposite current answers.
- The study uses 130 extractor-selected atomic transitions derived from software fixes; without a re-read, identity-based temporal memory reached 98.5% model-judged accuracy with zero observed stale-value errors.
- The verbatim re-read pushed the stale-value rate to 88.5%; the guard against reactivating retired values cut it to 0.8% in that constructed condition.
- The author's design takeaway is to separate observing a value from evidence that the value changed, and notes limits: selected inputs, aggregate-only answer records, related-family judges and a guard evaluated after the failure was found.
articleTemporal Validity on Real Software Histories: Eliminating Stale-Fact Errors in Code-Assistant Memory over GitHub FixesNeeraj Yadav
articleThe Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory AgentsJundong Hu, Shekar Ramachandran
articleWhen Evidence Changes: Evaluating Memory Repair and Re-reading in Language-Model AgentsWenhui Chu (University at Albany, State University of New York)
Checking sign-in…
Loading comments…