Ask an AI agent "how do I deploy to staging?" and you want an exact procedural match — the literal commands, in order. Ask "what were our database decisions?" and you want semantically related analytical records — the reasoning, the tradeoffs, the alternatives considered.
These are fundamentally different retrieval problems. But every existing memory system treats them identically: embed the query, find the nearest vectors, return results. That's leaving precision on the table.
Orion tags every knowledge record with a cognitive region that determines how it's stored, cached, retrieved, and reasoned about. The result: a 28% improvement in precision@5 compared to untyped retrieval.
The seven regions
| Region | Cache TTL | Retrieval bias | What it stores |
|---|---|---|---|
analytical |
8 hours | Semantic similarity | Decisions, tradeoffs, logical reasoning |
procedural |
24 hours | Keyword match | Commands, workflows, step-by-step instructions |
contextual |
4 hours | Recency | Project state, preferences, observations |
creative |
72 hours | Semantic similarity | Analogies, lateral connections, novel framings |
empathetic |
1 hour | Entity graph | Communication preferences, team dynamics |
critical |
8 hours | Keyword match | Failure patterns, edge cases, corrections |
strategic |
7 days | Semantic similarity | Goals, roadmaps, long-horizon plans |
The cache TTLs reflect how quickly each type of knowledge changes. Strategic knowledge stays hot for 7 days because quarterly goals don't shift daily. Empathetic context expires in 1 hour because relational dynamics are fluid — the communication preference that mattered in this morning's meeting may not apply this afternoon.
How regions change retrieval
Each region tunes the weights in Orion's RRF search pipeline:
Procedural queries get boosted keyword weight and a lower RRF k parameter (sharper ranking). When you search for "deploy to staging," you want the record that contains those exact words, not a semantically similar discussion about deployment philosophy.
Analytical queries get boosted semantic weight and a higher k (smoother blending). When you search for "database decisions," you want records about choosing Postgres, evaluating MySQL, and considering DynamoDB — even if none of them contain the word "decision."
Contextual queries get boosted recency weight. The current project state matters more than last month's state.
Empathetic queries get boosted graph weight. "How does Sarah prefer to communicate?" is best answered by traversing the entity graph from "Sarah" to connected stardust, not by semantic search.
ChromaDB partitioning
Each region gets its own ChromaDB collection per galaxy:
orion_{galaxy_id}_analytical
orion_{galaxy_id}_procedural
orion_{galaxy_id}_contextual
orion_{galaxy_id}_creative
orion_{galaxy_id}_empathetic
orion_{galaxy_id}_critical
orion_{galaxy_id}_strategic
This isn't just organizational — it enables region-specific embedding strategies. Procedural content benefits from different chunking than analytical content. We're exploring per-region embedding models in future versions.
Cognitive mode prompts
When brain.think or brain.recall specifies a cognitive_mode, Orion injects a region-specific reasoning prompt:
# analytical mode
"Draw on accumulated logical frameworks, decision records, and trade-off
analyses. Prioritize precision. Ask: what is the most defensible
conclusion given all evidence?"
# procedural mode
"Draw on established patterns, verified workflows, and step-by-step
procedures. Prioritize accuracy and completeness. Ask: what is the
exact sequence of steps?"
These prompts shape how the LLM reasons about the retrieved context, not just what context is retrieved. The combination of typed retrieval and typed reasoning produces noticeably better results than either alone.
Benchmark results
We evaluated on 500 query-answer pairs across 3 galaxies, comparing untyped retrieval (single ChromaDB collection, uniform RRF weights) against region-aware retrieval:
| Metric | Untyped | Region-aware | Δ |
|---|---|---|---|
| Precision@5 | 0.57 | 0.73 | +28% |
| Precision@1 | 0.64 | 0.79 | +23% |
| MRR | 0.61 | 0.76 | +25% |
The improvement is consistent across query types, but largest for procedural queries (+35%) where keyword matching dramatically outperforms semantic search.
The neuroscience parallel
The human brain doesn't store all memories in one place. Procedural memory (how to ride a bike) lives in the cerebellum and basal ganglia. Episodic memory (what happened yesterday) involves the hippocampus. Semantic memory (facts about the world) is distributed across the neocortex.
We didn't set out to model neuroscience — we set out to improve retrieval precision. But the parallel is instructive: the brain evolved separate memory systems because different types of knowledge have different access patterns. The same principle applies to AI memory.
Read the cognitive regions documentation or explore how regions interact with the RRF pipeline.