bdi-mental-states

Context is a finite budget; be lazy with tokens. The agent's head is the scarce slot, not the triple store. A belief is context that earned that slot — grounded in a world state, dated, justified. Everything else stays in triples. Don't promote a fact to a belief, a belief to a desire, or a desire to an intention until something downstream needs it. The smallest model that answers "why did the agent act?" is the right one.

The triad — load only what's load-bearing

State	Holds	Must reference	Promote when
Belief	what's true	a WorldState (never free text)	a fact is goal-relevant
Desire	what to bring about	the Belief that motivates it	a belief creates a goal
Intention	what's committed	the Desire it fulfils + a Plan	the agent commits to act

Endurants (states above, persist over intervals) vs perdurants (BeliefProcess, DesireProcess, IntentionProcess — events that create/modify states). States are nouns, processes are verbs — mixing them collapses the perception↔cognition boundary and breaks type-filtered queries.

Cognitive chain — the one standout pattern

Wire the triad with bidirectional pairs so the graph reads both directions: Belief —motivates→ Desire —fulfilledBy→ Intention —specifies→ Plan —hasComponent→ Task —precedes→ Task

Forward = "what should the agent do?" Backward = "why did it act?" The backward trace is the whole point: explainable agency for free. Properties: motivates/isMotivatedBy, fulfils/isFulfilledBy, isSupportedBy (intention←belief), specifies (intention→plan), addresses (plan→goal), precedes (task order), isJustifiedBy (any state→Justification), refersTo (belief→worldstate).

T2B2T — the bidirectional pipeline

Agents consume external RDF and emit new RDF. Two phases, provenance preserved:

1. Triples→Beliefs: WorldState —triggers→ BeliefProcess —generates→ Belief. Source data flows into cognition with a traceable origin.
2. (deliberate: belief→desire→intention→plan→execute)
3. Beliefs→Triples: PlanExecution —bringsAbout→ WorldState. Results project back as standard linked data so downstream systems consume agent output.

Build a BDI model — six passes

1. World-state substrate — model perceivable facts/events as WorldState first.
2. Beliefs — one per goal-relevant world state, with provenance + validity + justification.
3. Desires — only where a belief creates a motivation; link back to it.
4. Intentions — only on commitment; record the plan and preconditions.
5. Project back — emit result world states as RDF.
6. Validate — query provenance, motivation, plan order, active validity before trusting it.

Measured thresholds — don't overload the head

• Ontology: start with 9 classes (Agent, WorldState, Belief, Desire, Intention, Plan, Task, Justification, TimeInterval). Add a class only when a competency question proves the core can't answer it.
• Chain depth: ≤ 3 levels (belief→desire→intention). Deeper rarely improves decisions and explodes LLM inference cost.
• Every state gets a validity TimeInterval — beliefs without bounds can't be GC'd or conflict-checked.
• Composite belief → hasPart sub-beliefs, so a partial update doesn't replace the whole.

Logic Augmented Generation (folded algorithm)

Constrain LLM triple generation: serialize the ontology into the prompt, generate, extract, validate against class restrictions, retry-with-feedback on violation — the ontology in context is the guardrail.

Lazy is not negligent

Never skip: world-state grounding (ungrounded beliefs break querying and cross-agent interop), justifications (unjustified states can't be audited — the trace dies), temporal bounds (no GC, no conflict detection). These are the spine of explainability; cutting them isn't lean, it's silent drift.

Validate — competency questions

Ship only if these resolve: which beliefs motivated a desire? which desire does an intention fulfil? which process generated a belief? ordered task sequence? states valid at time T? No answer → undermodeled.

Boundaries

• Persistent facts / entity memory without BDI semantics → memory-systems.
• Agent topology, handoff, coordination → multi-agent-patterns.
• Quality gates / regression on BDI output → evaluation.
• Context-window + attention mechanics → context-fundamentals.
• Query/validation/projection tool contracts → tool-design.

"stop ponytail" / "normal mode": revert to verbose BDI mental-state guidance.