Relational cognition
Relational Belief States
A falsification-first investigation into whether persistent, directional, agent-local models of other agents provide measurable value.
Why it matters
Artificial agents increasingly act across repeated interactions, but we do not yet know what social continuity requires or when explicit relationship state is better than reconstructing context from history.
Current claim
A distinct RBS architecture is not uniquely necessary for the capabilities tested so far. Explicit relational state may still be valuable as an inspectable engineering interface under realistic memory, compute and auditability constraints.
Initial question
What must an agent persist about another agent for socially coherent behaviour to continue across time?
The initial intuition was that relationships have state: trust, commitments, expectations and uncertainty cannot always be recovered cheaply from a growing interaction history.
Initial hypothesis
I expected directional, agent-local relational state to provide an architectural advantage over generic memory. The minimal design kept one revisable record per ordered pair of agents, with provenance and uncertainty attached to updates.
Stronger baselines and state collisions
Early positive results were not enough. Generic-state baselines were made progressively stronger, given equivalent access to history, and prompted to reconstruct socially relevant context. Collision cases were added where a single undifferentiated state had to represent conflicting relationships.
Some bounded advantages survived narrow tests, especially around inspectability and directional ownership. The broader architecture claim did not.
Comparator escalation
Once the comparator could maintain arbitrary persistent state and retrieve the same evidence, much of the apparent RBS-specific advantage disappeared. That result narrows the claim: the value may come from explicit representation and operational constraints, not from a uniquely necessary cognitive architecture.
What survived falsification?
Explicit relational state remains a useful engineering interface for inspectability, directional ownership, modular retrieval, uncertainty, provenance, serialisation, replay and auditability.
The live research programme now asks when those properties improve capability-cost tradeoffs under realistic memory constraints, and how any useful representation should scale from relationships to groups and institutions.
Current direction
The next phase separates three questions that were previously bundled together: what information must persist, what representation makes that information usable, and what architecture maintains or retrieves it efficiently.
RBS is now one mechanism inside a wider programme on social intelligence, not the identity of the programme itself.
Evidence ledger
What the claim currently rests on
+ Supports
Inspectable state
Directional records make social updates easier to inspect, replay and audit than latent or reconstructed context.
Bounded collision cases
Explicit ownership helps in tests where agents hold conflicting relationship-specific information.
− Weakens
Stronger generic state
A capable generic-state comparator reproduced much of the behavioural advantage attributed to RBS.
Architecture dependence
Positive results were sensitive to comparator strength, weakening the general distinct-architecture thesis.
? Unknown
Long-horizon rarity
Whether explicit state preserves rare but important interactions better under strict memory budgets.
Beyond dyads
Whether group and institutional state require representations that cannot be reduced to pairwise relationships.
What would change my mind?
A retrieval-only or generic-state system matching or exceeding persistent relational-state agents across long-horizon social tasks under equal context, compute and memory budgets would remove most of the remaining capability case. Conversely, stable advantages under those controls would strengthen it.
Next steps
- Match context, memory and compute budgets more strictly.
- Test rare-event retention and delayed social consequences.
- Separate representation quality from update architecture.
- Build group- and institution-level tasks before extending pairwise state by assumption.
Research pages preserve changed claims rather than silently rewriting them. Last revised .