Temporal Mind: Experimental Cognitive Agent Framework

Motivation

Following the Stanford Smallville experiment and emerging work on agent frameworks, this project investigates how temporal awareness and memory retrieval patterns influence agent behavior over extended interactions. It is a testbed for exploring questions about:

Temporal Persistence: Does recency bias lead to stable behavioral patterns or continuous drift?

Emotional Dynamics: How does emotional state propagate through the perception-action loop?

Memory Formation: What patterns emerge in which experiences become memories?

Path Dependence: How do early experiences shape later behavior through memory accumulation?

Research Context

Inspired by:

• Park et al.'s Generative Agents (Stanford "Smallville" simulation) - demonstrating emergent social behaviors from memory-augmented LLM agents
• LangChain's agent framework
• Cognitive architectures for LLMs

Architecture

The system implements a simple perception-action loop with temporal awareness:

Clock (discrete time) → State (world observations) 
  ↓
Memory Retrieval (recency-based) → Bias Computation (emotional state)
  ↓
Action Generation → Memory Formation
  ↓
State Update → Clock Tick → [repeat]

Core Components

Clock: Discrete time tracker enabling temporal reasoning and memory recency calculations

State: Observable world conditions that update with agent actions and environmental changes

Memory: Timestamped experience records forming the agent's episodic history

MemoryGraph: Storage and retrieval system using temporal heuristics to surface relevant memories

BiasModel: Computes emotional biases from current observations and retrieved memories using LLM inference

CognitiveModel: Generates actions and forms new memories based on world state, emotional bias, and recent actions

FileSystem: Persistent storage for simulation histories enabling post-hoc analysis

Installation

# Clone repository
git clone https://github.com/yourusername/temporal-mind.git
cd temporal-mind

# Set OpenAI API key
export OPENAI_API_KEY='your-api-key-here'

# Install dependencies
pip install openai

Usage

cd src
python main.py

Simulation runs for 100 epochs (configurable in main.py). Histories save to timestamped directories in temporal_mind_data/ for analysis.

Configuration

Edit main.py to modify:

• DEFAULT_WORLD_STATE_OBSERVABLES: Initial environmental conditions
• DEFAULT_MEMORIES: Foundational knowledge/experiences
• ROUNDS: Number of time steps to simulate
• Memory retrieval parameter k in get_top_k_memories()

Example Output

temporal_mind_data/
└── 20250119/
    └── histories/
        └── histories_20250119_143022/
            ├── world_states.txt      # State evolution over time
            ├── biases.txt             # Emotional trajectory
            ├── reactions.txt          # Action sequence
            └── new_memories.txt       # Memory formation pattern

Observations and Limitations

Current State

The implementation successfully demonstrates:

• ✅ Stable perception-action loops over 100+ time steps
• ✅ Coherent memory formation and retrieval
• ✅ Emotional bias computation from memory and state
• ✅ Persistent state evolution with action consequences

Known Issues

• Memory graph grows unbounded (no forgetting mechanism)
• No memory consolidation across time scales
• Purely reactive (no planning or goal-directed behavior)
• Limited evaluation metrics

Project Structure

temporal-mind/
├── src/
│   ├── main.py              # Simulation loop and configuration
│   ├── clock.py             # Discrete time tracking
│   ├── state.py             # World state representation
│   ├── memory.py            # Memory objects with timestamps
│   ├── memory_graph.py      # Memory storage and retrieval
│   ├── model.py             # BiasModel and CognitiveModel (LLM inference)
│   └── filesystem.py        # Data persistence
├── LICENSE
└── README.md

Citation

Temporal Mind: Experimental Cognitive Agent Framework
https://github.com/one-2/temporal-mind
2025