- Who it helps
- Researchers testing persistent agent memory
- Availability
- Public October 18 preprint. The study describes synthetic rules, fictional entities and exactly traceable intermediate states.
Remembering events is only part of the task
A task might become complete only after two separate approvals. Retrieving one approval does not establish the derived status. The benchmark tests these dependencies across ten scenarios.
Treat reported failures as bounded evidence
The authors report missed stored facts and failures to propagate implied changes across four memory systems. Synthetic examples help inspect causality, but do not prove absence of training contamination or reliability in real work.
What you can do
- Read the evaluation setup and its limitations before applying the result. Open reference ↗
Where the evidence stops
Abstract read; dataset, retrieval systems and reported comparisons not executed.
Sources read. Public text checked November 15, 2026 UTC. No installation, model inference or full workflow executed.
Sources & further reading
Paper · paper
MemDerive memory evaluation ↗
October 18 abstract read; synthetic event streams and derived-state evaluation are author results.
Sen Gao and coauthors · page checked 2026-11-15
Coverage note: Abstract read; dataset, retrieval systems and reported comparisons not executed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
MemDerive evaluates whether a memory system can recover facts derived from a stream of events.