Cognition
How iCog reasons over your memory.
iCog is structured around a multi-stage cognition pipeline. LLMs are infrastructure, not the algorithm — they render structured outputs into language. The intelligence lives in the routing, memory model, and relationship graph between memories.
Pipeline stages
- Moment assessment — fast heuristics decide if a turn is trivial (greeting, command) or requires deep cognition.
- Recall — multi-stage retrieval over the three-layer ocean (surface / middle / deep) with HNSW + BM25 + RRF reranking.
- Context build — 368-dim context signatures with LLM compression.
- Reasoning — five thinking engines (pattern, contradiction, prediction, insight, relationship).
- Reply synthesis — final LLM call given the structured context.
- Learning — fire-and-forget storage of memory deltas and feedback.
Why this structure matters
Swapping one LLM for another should change output quality, not system behaviour. Removing the LLM from any one stage should leave the rest of the pipeline coherent — the structure is the algorithm.