Three-layer memory

Multi-agent orchestration

Memory combines three layers. A knowledge wiki with a structured schema, a hybrid recall that merges vector search via Qdrant and graph search via FalkorDB, and per-project memory. A model-based visual index adds recall over documents and videos.

Key points: Knowledge wiki · Hybrid vector and graph recall · Per-project memory · Visual index · Consolidation · Sovereignty.

Knowledge wiki

The first layer is a wiki of Markdown pages with a structured, versionable schema. It holds durable domain facts: customers, products, and technical configurations.

Hybrid vector and graph recall

The second layer merges vector search via Qdrant with the relationship graph via FalkorDB. A reranker re-ranks the results so only the relevant ones surface.

Per-project memory

A third layer keeps facts specific to each project and loads them at session startup, giving the agent useful context without overload.

Visual index

A vision model indexes documents and videos. Recall becomes multimodal and retrieves a page or a sequence from its visual content.

Consolidation

A periodic pass deduplicates and synthesises entries through a model. Memory stays clean, free of duplicates or accumulated contradictions.

Sovereignty

The embedder and the reranker run locally on GPU. No memory data leaves the infrastructure.