VoidOrigin // AI Division
itremembers.itcorrects.itevolves.
A cognitive substrate is waking up. Cognitive memory for machines, built from the research up.
join the waitlist↓>what it is.
We are building a memory model that decays, corrects, anticipates, and earns its own identity. Not another retrieval layer. A cognitive substrate designed from the research up.
Fig. 01Live system metrics
Every coding agent has the same failure mode: the context was right there and it ignored it. The information existed in a prior session, in a correction the user already made, in a concept the agent retrieved but ranked below the cutoff. The context window is not memory. It is a buffer that empties between sessions. Retrieval-augmented generation gets halfway. It stores text and fetches by similarity, but similarity is not understanding. A system that retrieves without decaying, correcting, or anticipating is a search engine with a language model on top.
We are building something different. At the core is a five-pillar graph database on SurrealDB: agents, projects, tasks, artifacts, and concepts, connected by typed semantic edges (broader, narrower, supersedes, caused_by, supports, contradicts). Embeddings come from BGE-M3 at 1024 dimensions, generated locally with zero API dependency, indexed via HNSW for sub-second vector search across the full graph. A learned cross-attention reranker (ACAN, roughly 130K parameters) trains itself on your retrieval outcomes: every time a recalled memory is used or ignored, the signal feeds back into the model via pure TypeScript SGD. Training takes about 30 seconds. The whole system runs as a two-process architecture: a long-lived daemon holds the graph, embedding model, and reranker weights in a single memory pool; thin per-session clients forward calls over Unix socket IPC. Multiple sessions share one daemon. No hosted vector DB. No retrieval API. Nothing leaves the machine on the read path.
Core mechanisms
Stability decay
Every concept loses retrieval priority over time unless recalled or reinforced. Continuous decay modeled on Kanerva and Tulving, weighted by access frequency and edge density, with a floor at 0.15. Nothing disappears entirely.
Supersede edges
Corrections never overwrite. The system writes a new belief, links it to the old one with a typed edge, and decays the stale node to 40% stability. Both remain queryable. Correction is transition, not erasure.
Gap detection
Prospective recall traverses graph neighbors beyond the similarity top-N and surfaces what the vector search missed. It answers “what am I forgetting?” before anyone asks.
Soul graduation
After enough sessions with sufficient quality signals, the agent earns an identity document grounded in evidence from its own history. Identity is not configured. It is graduated.
98.2% Recall@5 on LongMemEval. 1740 tests passing, with schema-edge integrity guards that validate every graph relation against declared types at merge time. The retrieval pipeline generates 128 vector candidates, reranks through cross-attention, and returns grounded results without a single external API call. Every invariant is enforced, not hoped for. This is the closest thing to real cognitive memory in an AI model, and we are releasing it as a standalone system.
Early access
Join the waitlist.
We are releasing this as a standalone system. Drop your email and we will reach out when it is ready.