Loomcycle now ships agentic memory as a first-class runtime capability. One per-scope memory holds three kinds of content (facts, notes, documents) reached through one search API with honest source selectors. A background consolidator extracts ontology-typed entity facts from finished chat transcripts and inherits compaction-discarded spans so nothing useful is thrown away. Bi-temporal validity means a correction never destroys the record. Documents embed on write (prose, Mermaid across ten dialects, images via persisted vision descriptions) so every chunk is findable by content. Inline links become graph edges, transclusion works at export, backlinks/related/unlinked_mentions answer discovery in one call, and per-chunk history/diff is built in. Hybrid vector + full-text retrieval with Reciprocal Rank Fusion; narrowing to facts on a 3,000-chunk scope is 31x faster, not slower. Every write carries provenance (tenant, session, run, agent, model, compaction generation). A human-facing operator console (@loomcycle/memory-view) and GDPR-shaped subject erasure on every transport complete the shape.
What I Learned Stealing Ideas from Matt Pocock’s `.agents` Directory