Cognitive regions

Seven types of knowledge, each with tuned caching, retrieval weights, and reasoning prompts.

Every stardust record is tagged with a cognitive region. This isn't metadata — it's a first-class dimension that changes how the record is cached, retrieved, and reasoned about.

The core insight: different types of knowledge have different access patterns. A procedure ("how to deploy") needs exact keyword matching. A decision ("why we chose Postgres") needs semantic similarity. A goal ("reduce latency by 50%") needs to persist in cache for days. Treating them identically leaves retrieval precision on the table.

The seven regions

Region	Cache TTL	Retrieval bias	Stores
analytical	8h	Semantic similarity	Decisions, tradeoffs, reasoning
procedural	24h	Keyword match	Commands, workflows, how-tos
contextual	4h	Recency	Project state, preferences
creative	72h	Semantic similarity	Analogies, lateral connections
empathetic	1h	Entity graph	Communication preferences, dynamics
critical	8h	Keyword match	Failure patterns, edge cases
strategic	7d	Semantic similarity	Goals, roadmaps, priorities

How regions affect retrieval

Orion's search pipeline uses Reciprocal Rank Fusion to blend three signals: keyword cache, vector similarity, and graph traversal. Regions tune the weights:

• Procedural queries boost keyword weight and lower the RRF k parameter (sharper ranking). You want the exact deploy command, not a paraphrase.
• Analytical queries boost semantic weight and raise k (smoother blending). You want related decisions even if they use different words.
• Contextual queries boost recency. Current state matters more than last month's state.
• Empathetic queries boost graph weight. "How does Sarah communicate?" is best answered by traversing from the "Sarah" entity.

In benchmarks, region-aware retrieval improves precision@5 by 28% compared to untyped retrieval.

Storage partitioning

Each region gets its own ChromaDB collection per galaxy:

orion_{galaxy_id}_analytical
orion_{galaxy_id}_procedural
orion_{galaxy_id}_contextual
...

This enables region-specific embedding strategies and prevents procedural content (short, command-heavy) from polluting the vector space of analytical content (long, reasoning-heavy).

Cognitive mode prompts

When brain.think or brain.recall specifies a cognitive_mode, Orion injects a region-specific reasoning prompt that shapes how the LLM processes retrieved context:

# analytical mode
"Draw on accumulated logical frameworks, decision records, and trade-off
 analyses. Prioritize precision. What is the most defensible conclusion?"

# procedural mode
"Draw on established patterns, verified workflows, and step-by-step
 procedures. Prioritize accuracy and completeness. What is the exact
 sequence of steps?"

The combination of typed retrieval and typed reasoning produces better results than either alone.