{"slug": "the-loomcycle-memory-architecture-drawn", "title": "The loomcycle memory architecture, drawn", "summary": "The loomcycle memory architecture, as of v1.75.0, spans five storage planes—key/value, vector, SQL Memory, chunked-graph Documents, and an entity graph—with four scopes (agent, user, tenant, global) and server-side owner resolution to prevent cross-tenant data access. Facts carry three timestamps plus created_at, enabling as-of queries and superseding corrections, while hybrid recall uses Reciprocal Rank Fusion and a provenance envelope tracks every row. The system fails closed on placement uncertainties and includes an RFC CT P2 harvest bridge to convert distillation-evicted spans into durable facts.", "body_md": "Companion to the memory testing story. A structure-diagram tour of the memory subsystem as it stands after v1.75.0. One Memory tool reaches five storage planes: key/value (semantic recall), vector (embedding-backed nearest-neighbor), SQL Memory (per-scope relational store), chunked-graph Documents (hierarchical Markdown with typed chunks), and an entity graph layered on the Document chunks that carries facts with subjects, types, edges, and bi-temporal validity. Four scopes name whose data an op reaches: agent, user, tenant, global (admin-only, fails closed). Owner id is resolved server-side from the run identity, never wire-supplied; a model-supplied owner id would let one tenant's agent read another's data. A fact carries three timestamps (observed_at, valid_at, invalid_at) plus a stored created_at; the as_of predicate walks the timeline at any point, and a correction supersedes rather than deletes, so what was true on a past date still answers correctly and a fact with a known future end date does not disappear today. Ontology-declared placement decides which scope a fact about a given entity type belongs in; the decision tree fails closed on every uncertainty (undeclared type, inconsistently-typed subject, subject matches profile owner, isolated caller, missing grants). Both halves of a fact land in the same scope or neither does. Hybrid recall combines a vector search and a full-text search with Reciprocal Rank Fusion, filters by observed_at window, as_of, source selector, and tenant fold. A sources_applied flag whose zero value is false makes a silently-widened result impossible to pass as filtered. Every row carries a provenance envelope: tenant, session, run, agent, model, compaction_generation, source_quote span, origin (server-stamped from the writer's identity), judged_by, judged_at. An RFC CT P2 harvest bridge banks distillation-evicted spans onto the consolidator's queue so a span the run would have thrown away becomes a durable fact for future runs. Six Mermaid structure diagrams.", "url": "https://wpnews.pro/news/the-loomcycle-memory-architecture-drawn", "canonical_source": "https://loomcycle.dev/blog/loomcycle-memory-architecture.html", "published_at": "2026-09-03 14:00:00+00:00", "updated_at": "2026-09-03 15:56:41.855649+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-research"], "entities": ["loomcycle"], "alternates": {"html": "https://wpnews.pro/news/the-loomcycle-memory-architecture-drawn", "markdown": "https://wpnews.pro/news/the-loomcycle-memory-architecture-drawn.md", "text": "https://wpnews.pro/news/the-loomcycle-memory-architecture-drawn.txt", "jsonld": "https://wpnews.pro/news/the-loomcycle-memory-architecture-drawn.jsonld"}}