Skip to content

Cognitive Architectures

Maturity: Experimental

All cognitive architecture components are marked @component_maturity("experimental"). They implement simplified versions of classical architectures for research and exploration. They have not been validated against standard cognitive science benchmarks and should not be used for production decision-making.

Six components implementing simplified classical and modern cognitive architectures for structured problem-solving research.

Overview

Cognitive Architectures provides G6 with multiple reasoning strategies inspired by cognitive science research. These are simplified implementations suitable for exploring architectural patterns, not full-fidelity replicas of the original systems.

cog_arch_gps implements means-ends analysis from the General Problem Solver tradition, reducing the difference between current and goal states through operator selection. cog_arch_actr provides a simplified ACT-R production system with procedural and declarative memory. cog_arch_soar brings production rules and chunking mechanisms. cog_arch_aixi and cog_arch_dgm are theoretical/research-oriented implementations. deep_understanding provides cross-domain conceptual reasoning.

Components

Component Description MCP Tools
cog_arch_actr ACT-R production system (procedural/declarative memory, activation-based retrieval) --
cog_arch_aixi AIXI-inspired bounded CTW/MCTS planning; not a validated reliability guarantee --
cog_arch_dgm Deep generative model architecture (latent space modelling, generative reasoning) --
cog_arch_gps General Problem Solver (means-ends analysis, operator selection) --
cog_arch_soar Soar cognitive architecture (production rules, chunking, impasses) --
deep_understanding Deep conceptual understanding (cross-domain reasoning) --

Architecture

graph TD
    ACTR[cog_arch_actr] --> DU[deep_understanding]
    SOAR[cog_arch_soar] --> DU
    GPS[cog_arch_gps] --> DU
    AIXI[cog_arch_aixi] --> DU
    DGM[cog_arch_dgm] --> DU
    GPS -->|means-ends analysis| SOAR
    ACTR -->|declarative memory| DGM
    AIXI -->|candidate policy| GPS
    DU -.->|cross-domain transfer| EXP[Experience & Autonomy]
    DU -.->|reasoning strategies| GOAL[Goal & Planning]

Key Patterns

Production Systems. Both ACT-R and Soar are production-rule architectures, but they differ in conflict resolution. ACT-R selects productions by activation level (a continuous utility score decaying over time), while Soar fires all matching productions in parallel and resolves conflicts through impasse-driven sub-goaling. G6 can route a task to whichever strategy fits: ACT-R for memory-intensive recall tasks, Soar for problems requiring iterative decomposition.

Means-Ends Analysis. GPS applies the classical reduce-the-difference strategy: identify the gap between current state and goal state, select an operator that reduces that gap, and recurse. This pairs naturally with goal_engine decomposition -- GPS provides the operator-selection logic within each sub-goal.

Universal Prediction. The AIXI component uses CTW-style sequence prediction and MCTS lookahead to recommend actions from observed action/observation/reward history. Its MCP implementation is built for durable local use, with SQLite persistence, restart recovery, bounded inputs, and explicit human-gate semantics. The algorithmic approximation is still experimental: action recommendations are hypotheses to test, not guarantees of optimality, safety, or real-world reliability.

Generative Reasoning. The DGM component models problems in a learned latent space, enabling interpolation between known solutions and generation of novel candidates. This complements the symbolic approaches by handling continuous, high-dimensional problem spaces where discrete production rules struggle.