EvoWiki Replaces Top-k Retrieval With Entity Version Chains and a State-Overwrite Protocol for Evolving Knowledge

arXiv 2608.23265·medium signalXBlueskyLinkedInCopy link

In collaboration spanning multiple meetings, facts like decisions and risks get revised and overturned, but long-context methods stack the whole history while RAG and structured-memory methods store facts as static or append-only and rank by semantic relevance at read time, so stale and current states get retrieved together. EvoWiki separates offline incremental construction from online structured reading, capturing intra-meeting evolution from proposal to decision and using entity version chains plus a fine-grained State-Overwrite Protocol to distinguish currently valid state from superseded history while keeping meeting-level provenance anchors. The read path bypasses relevance-based top-k entirely in favor of deterministic entity addressing, temporal resolution, and cross-entity multi-hop aggregation.