The architecture that lets twenty companies share one organizational mind.
Three kinds of memory, one knowledge graph.
Semantic memory holds facts. Episodic memory holds what happened, when, and to whom. Procedural memory holds reusable patterns extracted from successful episodes. Every agent reads and writes all three.
Bayesian confidence, with decay.
Agents track confidence in their own knowledge through Bayesian inference. Confidence decays without reinforcement. When confidence drops below threshold, agents hedge or escalate — they don't bluff.
The right memory at the right level.
Memories are attributed to one of six scopes — this user, this agent, this task, this organization, this domain, or universal. Learning at one level can elevate to others when patterns repeat.
Vector + BM25 + Graph, fused with RRF.
Pure vector search misses the obvious. Pure keyword search misses the nuanced. Memriq combines vector embeddings, BM25, and knowledge-graph traversal — fused via Reciprocal Rank Fusion — so agents find the right context even with vague queries.
Every decision is traceable.
The knowledge graph stores all memory, relationships, and ontology. Every agent decision links back to the facts and episodes that informed it. Auditability is structural, not bolted on.
Memriq's memory architecture draws on AriGraph, MemGPT, HippoRAG, GraphRAG, MACLA, and the CoALA framework. We didn't pick one — we built a hybrid that takes what each does best.