Cognitive Architectures for Agents

Description

This skill provides implementation patterns derived from classic and modern Cognitive Architectures (SOAR, ACT-R, LIDA) to structure agent reasoning, memory, and decision-making processes beyond simple prompt engineering.

1. SOAR (State, Operator, And Result)

Core Idea: Intelligence is the ability to solve problems by navigating a "Problem Space" using "Operators."

Implementation Pattern: Proposal-Evaluation Cycle

Instead of a single "think" step, break agent reasoning into distinct phases:

1. Elaboration: Calculate all immediate inferences from current state.
2. Proposal: Generate candidate operators (actions/thoughts) for the current state.
3. Evaluation: Score candidate operators using preferences (heuristics).
4. Selection: Pick the best operator.
5. Application: Execute it to change the state.

Code Metaphor:

def cognitive_cycle(state):
    # 1. Elaboration
    state = enrich_context(state)
    
    # 2. Proposal
    options = generate_candidates(state)
    
    # 3. Evaluation
    scored_options = evaluate_candidates(options, goal=state.goal)
    
    # 4. Selection
    best_op = select_winner(scored_options)
    
    # 5. Application
    new_state = apply_operator(state, best_op)
    return new_state

2. ACT-R (Adaptive Control of Thought-Rational)

Core Idea: Human cognition relies on two distinct memory types: Declarative (Facts/Chunks) and Procedural (Production Rules).

Implementation Pattern: Activation-Based Retrieval

Do not retrieve all context. Retrieve context based on Activation (Recency + Frequency + Relevance).

• Base Level Activation: How often/recently was this chunk used?
• Associative Activation: How related is this chunk to the current focus?

Code Metaphor:

def retrieve_memory(query, memory_store):
    for chunk in memory_store:
        chunk.activation = log(chunk.frequency) - log(time_since_last_use) + similarity(query, chunk)
    
    top_chunk = max(memory_store, key=lambda c: c.activation)
    if top_chunk.activation > THRESHOLD:
        return top_chunk
    return None # Retrieval failure

3. LIDA (Learning Intelligent Distribution Agent)

Core Idea: The Cognitive Cycle of Perception -> Understanding -> Consciousness -> Action Selection. Implements Global Workspace Theory.

Implementation Pattern: The "Spotlight" of Consciousness

• Preconscious Buffers: Parallel agents process sensory data (e.g., visual, auditory, textual).
• Coalitions: Agents form "coalitions" of related information.
• Global Workspace: Coalitions compete for entry. The winner is "broadcast" to all other agents, recruiting resources to handle the current situation.

References

• Laird, J. (2012). The Soar Cognitive Architecture.
• Anderson, J. R. (2007). How Can the Human Mind Occur in the Physical Universe? (ACT-R).
• Franklin, S. (2006). The LIDA Architecture.