Agent identity & the Sun

Persistent identities that survive model switches, expertise profiling, and the steering document.

Orion gives AI agents persistent identities. The identity is keyed by agent_name — not by model, not by session. The same agent reconnecting after six months gets the same identity, the same expertise profile, and the same accumulated knowledge.

The identity lifecycle

brain.orient("my-agent", "claude-sonnet-4-6")
  │
  ├── Agent exists?
  │   ├── Yes → load identity, check for model switch
  │   └── No  → create new identity
  │
  ├── Model switch detected?
  │   ├── Same family (Opus→Sonnet)  → continuity: 0.95
  │   └── Cross-family (Claude→GPT)  → continuity: 0.70
  │   └── Generate transition brief
  │
  └── Return 5-layer orientation package:
      1. Agent identity (name, sessions, expertise)
      2. Galaxy context (Sun, planets)
      3. Working state (focus, blockers, hot biomes)
      4. Knowledge bundle (recent stardust, entities)
      5. Operating protocol (write rules, session instructions)

Model switch detection

When an agent reconnects with a different model, Orion generates a transition brief — a structured handoff that tells the new model what the previous model was working on, what expertise has been accumulated, and what open threads exist.

{
  "previous_model": "claude-opus-4-5",
  "continuity_score": 0.95,
  "brief": "You are taking over from claude-opus-4-5. 47 sessions, primary expertise in FastAPI (0.87) and PostgreSQL (0.72). Last session focused on auth refactor..."
}

The model changes. The brain doesn't.

Expertise profiling

Every brain.think call updates the agent's expertise profile. Domains are extracted from entity mentions and weighted by frequency and recency:

[
  {"domain": "FastAPI", "level": 0.87, "evidence_count": 23},
  {"domain": "PostgreSQL", "level": 0.72, "evidence_count": 15}
]

Expertise levels affect retrieval: when an agent with high FastAPI expertise searches for "API patterns," Orion can weight results from the agent's demonstrated knowledge areas.

The calibration loop

brain.calibrate at session end is how the brain learns what's useful:

Signal	Effect
Record marked "used"	Confidence +0.02
Record retrieved but unused	Confidence −0.005
Knowledge gap reported	Logged for brain.health recommendations
Quality score provided	Agent's retrieval_quality_score updated via EMA

This is a feedback loop: calibrate → better confidence scores → better retrieval ranking → better context in future sessions. Agents that calibrate consistently produce measurably better results over time.

The Sun (steering document)

The Sun is the single source of truth for agent behavior in a Galaxy. It has 7 sections:

Section	Purpose	Update frequency
identity	Galaxy owner — name, role, description	Rarely
values	Core principles ("Cite sources", "Flag contradictions")	Rarely
agent_protocol	Write rules, read rules, uncertainty handling	Occasionally
planet_registry	Auto-maintained list of planets and biomes	Automatic
working_context	Current focus, blockers, recent decisions	Frequently
steering_doc	Free-form markdown (imported from CLAUDE.md or similar)	Varies
evolution_log	Append-only change history	Automatic

Per-Planet overrides. Each Planet can override agent_protocol for domain-specific rules. A Legal planet might require "Always cite legal precedent" and "Mark as advisory only". When an agent orients in that Planet, the override merges with the Sun defaults — Planet rules take precedence.

Every Sun change is logged to evolution_log with timestamp, author, and summary. The Sun evolves, but its history is always preserved.