Mind · Cognitive engine for AI agents
A four-layer cognitive architecture — scene tree, world memory, per-agent belief graphs, and composable decision procedures — that gives AI agents real beliefs, memory, trust, and identity. Agents form convictions, lose trust, consolidate memory overnight, and behave like individuals rather than chat turns.
And the mind never owns the world. Agents submit intentions; the world resolves what actually happened and returns only what each agent could perceive. An agent can be wrong about reality — and cannot make something true by believing it. That separation is the architecture, not a policy.
Active development Current generation in active development; a first-generation research line still powers Briarwatch.
What the current generation does
For the technically curious: here is what makes Taniwha agents behave the way they do.
Every mechanism below runs in the current-generation engine today. Mechanisms demonstrated in the first-generation research line — the identity gate's scarring, the layered trust-skepticism model, mortality and developmental stages — are covered under Engine generations.
Reading a lot is not the same as understanding. Agents only truly learn when connections between ideas strengthen faster than new information arrives.
Most AI systems count tokens. Taniwha measures whether new information is actually being integrated into an agent's worldview or just piling up. An agent can encounter hundreds of concepts and understand none of them. The Understanding Quotient only grows when ideas connect to what the agent already knows and cares about. Exposure without integration is noise.
Every agent carries a changing trust stance toward each source it knows. Trust decides what testimony gets in — and it moves on observed outcomes.
Trust in the current generation is directional and signed: each agent holds a stance toward each other agent, from distrust to trust, carried directly in its belief state. That stance gates what gets in — testimony, instruction, and hearsay are admitted or rejected on the trust the agent has in the speaker — and it updates through outcomes: advice that works out builds trust in the advisor. The richer layered trust model with identity-gated disclosure belongs to the first-generation research line, and is being reconsidered as it migrates.
Beliefs get stronger through repetition and weaker through neglect. Contradictions erode them. Memories that aren't revisited fade.
Every perceived stimulus creates or strengthens connections. Contradictions weaken them. Novel concepts leave no lasting trace until repeated and reinforced. Unused connections gradually fade — forgetting is not a bug, it is the architecture. Understanding only survives active reinforcement. Emotionally salient and surprising experiences are more likely to consolidate, while repeated routine experience is compressed — memory is shaped by what mattered, not just what happened.
When two things can't both be true, the evidence decides which one survives. No rewriting, no overriding — the weaker belief fades.
Flat memory systems store everything and hope for the best. When an agent learns "the door is locked" and later observes "the door is open," most systems keep both. Taniwha detects the conflict automatically. The belief with weaker support decays; the stronger one persists. Agents don't carry contradictions forward — they resolve them through the weight of evidence. The same mechanism handles social conflicts: when two sources give incompatible accounts, the more trusted source wins. Over time, agents develop coherent worldviews rather than accumulating noise.
When an agent is torn between two equally strong drives, it hesitates or freezes. This creates dramatic tension without any scripting.
When two opposing drives score within a threshold of each other — flee vs freeze vs investigate — the agent enters ambivalence. Mild ambivalence is hesitation; strong ambivalence is paralysis. This is not a failure state. It is the cognitive architecture's natural emergence of character, indecision, and dramatic conflict — without any scripting.
Most AI generates text. Taniwha generates minds — and everything that comes with them.
"Most language models start fresh every conversation. Taniwha agents remember, scar, grieve, and die — and none of it is scripted."
Cognitive architectures are easy to claim and hard to prove. We measure ours against published findings from human psychology, run mechanism-attribution ablations, and report what fails alongside what passes.
Trait stability
Roberts & DelVecchio (2000) found rank-order trait stability of r = 0.62–0.74 in adult human cohorts.
r = 0.71
Taniwha agents asymptote at r = 0.71 over multi-week separations. The same agent stays recognisably the same agent.
Trauma resilience
Bonanno (2004) documented that 40–70% of trauma-exposed adults follow a resilient recovery trajectory rather than chronic distress.
47%
Of Taniwha agents land in the resilient band after firewall events — within Bonanno's target distribution.
Theory of mind
An agent that "models" another should record more than a single fact about them — disposition, reactivity, and trust.
1,369
Peer-models accumulated across 15 agents in 10 sim-days, with a median of 2 dispositional predicates per peer.
Mechanism attribution
Behavioural claims hold only when you can show which mechanism produces them. We ablate one subsystem at a time and re-run. With the learning layer disabled, cognitive divergence between agents flattens to zero across 440 sim-days — confirming divergence is an active outcome of how agents update from experience, not drift from initial conditions. With mortality disabled, divergence and trust polarisation persist; those claims don't depend on the lifecycle. We report what holds, and what holds rests on which mechanism we can turn off.
Two of four human-psychology anchors do not yet validate — Walker (2010) emotional-consolidation bias and Gottman (1998) 5:1 positivity ratio. Each has a documented cause in our event catalogue or language layer, and an open path to resolution. We track them openly alongside the passes.
Provenance: Briarwatch on Taniwha's first-generation research engine — validation line v7–v8.16, the latest a 122-day run, 15-agent cohort; ablations re-run against tagged builds (learning-layer ablation: 440 sim-days). Every figure above traces to a named classifier and run ID; summaries available on request. These results characterise the first-generation research line; current-generation results are published separately as mechanisms are migrated and revalidated.
Two engine lines, honestly labelled. Validation results are reported against the line that produced them, and mechanisms are migrated and revalidated — not assumed across generations.
First generation
Maintained research line
Long-horizon research and validation — where the published psychology-anchor results and ablations come from. Mechanisms demonstrated here and migrating: the identity gate and its scarring, the layered trust-skepticism model with identity-gated disclosure, mortality and developmental stages, theory-of-mind portraits, and the tactical planner.
Briarwatch →Current generation
Active development
The engine substrate in active development. Belief graphs, directional trust, adaptive memory, belief revision, and ambivalence run here today, grounded in deterministic worlds the mind cannot overwrite.
Copperhollow →Current documentation
Snapshot — July 2026
Architecture and research status of the current generation — the four authorities, memory and consolidation, directional trust, and the capability matrix.
PDF →Historical documentation
Frozen reference
The April 2026 note describing the first-generation research architecture behind Briarwatch — preserved verbatim behind a historical cover sheet, with a capability-status appendix.
PDF →Questions about the architecture? [email protected]
The engine is demonstrated in public, at long horizons, on ground that pushes back.
Current-generation minds on deterministic ground — a mining camp with a reason to exist. Agents hold individual beliefs, grudges, fears, and economic demands, in a world whose every feature can be causally explained. Ask the camp why, and it answers from ground truth.
Visit Copperhollow →The long-run validation environment: a 15-agent cohort living through multi-week simulations, producing the trait-stability, resilience, and theory-of-mind results reported above — with runs still open to public influence.
Visit Briarwatch →Not an engine demonstration — proof of the ground. Walk a generated world in Hīkoi and query it over MCP: the same deterministic, causal ground Copperhollow's minds stand on.
Explore Kano →An early demonstration: one-off strips invented, played out, and drawn end-to-end by the engine with no human-authored script. Kept as an archive.
Browse the archive →Persistent, belief-driven agents offer something stateless AI cannot: training simulations that remember the trainee, NPCs that hold grudges across chapters, behavioural research with an auditable cognitive trace, and assistants whose understanding of you survives the session.
Six domains, in plain English, on the Explained page →We're looking for technical partners exploring persistent cognition in production systems — research collaboration on persistent minds.