Beyond Code: Temporal Supersession Generalizes to World Knowledge
Oracle experiments on synthetic world-knowledge facts hit 1.000 accuracy / 0.000 stale in that protocol. Remaining error is extraction. This note is not the Paper 4 conversational 27B evaluation.
By Neeraj Yadav. MemStrata is built by Neeraj Yadav.
The breadth thesis
Earlier code-memory studies motivated this separate world-knowledge note. It asks whether deterministic supersession still holds on synthetic world-knowledge facts. We built a marker-free generator (CEO, population, price-style value flips on fictional entities to avoid real-world temporal ambiguity).
With LLM extraction (code-specialized 3B/7B models on prose), temporal_v6 reached 0.85 accuracy / 0.075 stale vs naive RAG 0.60/0.225 and advanced RAG 0.45/0.550. Better — but not zero residual.
Oracle isolation: residual is extraction, not mechanism
The decisive experiment builds the ledger from gold triples (no LLM extraction). Result on that oracle protocol: temporal_v6 1.000 accuracy / 0.000 stale. The residual vanishes when keys are correct. That isolates extraction error for this synthetic generator. It does not prove the mechanism generalizes to every domain or to live product use.
What comes next
The remaining work in this note is (1) a general instruction-model extractor for prose facts and (2) real suites (TEMPO, ECT-QA) using bi-temporal as_of(T) reads. Entity alias tables matter when USA and United States co-refer.
Do not read this as the conversational 27B evaluation, and do not treat oracle accuracy as a product guarantee.
Previous: Detect-and-Flag: Honest Memory for the 76% of Code That Is Not a Triple Next: Comparing MemStrata, Mem0, Graphiti and Letta: what the evidence can tell you