rotalabs-fieldmem

Field-Theoretic Memory System - Field-theoretic memory for AI agents with natural decay and consolidation.

Overview

FieldMem implements a physics-inspired approach to AI memory, treating memories as continuous fields that evolve according to partial differential equations. This enables:

• Natural forgetting: Memories decay over time following thermodynamic principles
• Semantic diffusion: Related memories influence each other spatially
• Importance-weighted retention: Important memories resist decay
• Collective intelligence: Multi-agent memory sharing through field coupling

Based on the heat equation:

∂u/∂t = α∇²u - γu + η(t)

Where:

• u: Memory field state
• α: Diffusion rate (memory spreading)
• γ: Decay rate (natural forgetting)
• η(t): Thermal noise (stochastic fluctuations)

Installation

pip install rotalabs-fieldmem

Optional dependencies

# For sentence-transformers embeddings
pip install rotalabs-fieldmem[embeddings]

# For visualization
pip install rotalabs-fieldmem[viz]

# For Google Cloud/Vertex AI integration
pip install rotalabs-fieldmem[gcp]

# All optional dependencies
pip install rotalabs-fieldmem[all]

Quick Start

Basic Memory Field

from rotalabs_fieldmem import MemoryField, FieldConfig

# Create a memory field
config = FieldConfig(
    shape=(128, 128),
    diffusion_rate=0.01,
    temperature=0.05,
)
field = MemoryField(config)

# Inject a memory
import jax.numpy as jnp
embedding = jnp.ones(64)  # Example embedding
field.inject_memory(embedding, position=(64, 64), importance=0.8)

# Evolve the field (memories diffuse and decay)
field.step()

# Query memories
values, positions = field.query_memories(embedding, k=5)

FTCS Agent

from rotalabs_fieldmem import FTCSAgent, AgentConfig

# Create an agent with field-theoretic memory
config = AgentConfig(
    memory_field_shape=(128, 128),
    diffusion_rate=0.003,
    temperature=0.05,
)
agent = FTCSAgent(agent_id="agent_1", config=config)

# Store memories
agent.store_memory("The capital of France is Paris.", importance=0.9)
agent.store_memory("Python is a programming language.", importance=0.7)

# Retrieve relevant memories
results = agent.retrieve_memories("What European capitals do you know?")
for memory in results:
    print(f"- {memory['content']} (score: {memory['combined_score']:.2f})")

Importance Scoring

from rotalabs_fieldmem import SemanticImportanceAnalyzer, QuickImportanceScorer

# Detailed semantic analysis
analyzer = SemanticImportanceAnalyzer()
scores = analyzer.analyze("URGENT: Meeting tomorrow at 3 PM!")
print(f"Total importance: {scores.total:.2f}")
print(f"Temporal score: {scores.temporal:.2f}")

# Quick scoring for real-time use
scorer = QuickImportanceScorer()
importance = scorer.compute_importance("Remember to call John Smith.")

Key Features

1. Field-Theoretic Memory

Memories are stored as energy distributions in a continuous field that evolves over time:

• Diffusion: Memories spread to nearby regions, creating semantic neighborhoods
• Decay: Low-importance memories fade naturally
• Noise: Thermal fluctuations add stochastic forgetting

2. Importance-Weighted Dynamics

The SemanticImportanceAnalyzer scores memories based on:

• Entity detection: Names, dates, locations
• Causal relationships: "because", "therefore", "leads to"
• Temporal markers: Deadlines, urgency, recurring events
• Instructional content: How-to, step-by-step procedures
• Emotional significance: Personal preferences, strong opinions

3. Multi-Agent Collective Intelligence

from rotalabs_fieldmem import MultiAgentSystem, MultiAgentConfig

# Create a multi-agent system
config = MultiAgentConfig(
    num_agents=4,
    consensus_threshold=0.6,
)
system = MultiAgentSystem(config)

# Agents share knowledge through field coupling
response = system.collective_query("What do we know about machine learning?")

4. Field Coupling Topologies

• Fully Connected: All agents share equally
• Nearest Neighbor: Spatial locality
• Small World: Local + random long-range connections
• Adaptive: Coupling strength based on semantic similarity

Architecture

src/rotalabs_fieldmem/
├── fields/           # Core field implementations
│   ├── memory_field.py   # PDE-based memory field
│   ├── sparse.py         # Sparse optimization (42x memory reduction)
│   └── operations.py     # Field operations
├── agents/           # FieldMem-enabled agents
│   ├── ftcs_agent.py     # Single agent
│   ├── multi_agent.py    # Multi-agent system
│   ├── coordinator.py    # Agent coordination
│   ├── collective.py     # Collective memory pool
│   └── coupling.py       # Field coupling
├── memory/           # Memory processing
│   ├── importance.py     # Semantic importance scoring
│   ├── consolidation.py  # Memory consolidation
│   └── forgetting.py     # Thermodynamic forgetting
├── utils/            # Utilities
│   ├── metrics.py        # Quality metrics
│   ├── embeddings.py     # Embedding management
│   └── persistence.py    # State persistence
└── viz/              # Visualization (optional)
    └── plotting.py       # Field visualization

Benchmark Results

Evaluated on standard long-context conversation benchmarks:

LongMemEval (ICLR 2025)

Memory Ability	FieldMem	Baseline	Improvement	Significance
Multi-session	0.042 F1	0.020 F1	+116%	p<0.01, d=3.06
Temporal reasoning	0.120 F1	0.084 F1	+43.8%	p<0.001, d=9.21
Preference tracking	0.268 F1	0.168 F1	+59.1%	p<0.001, d=8.96
Knowledge update	0.147 recall	0.115 recall	+27.8%	p<0.001

LoCoMo (ACL 2024)

Category	FieldMem	Baseline	Improvement
Category 3 (F1)	0.017	0.016	+9.5%* (p<0.05)
Category 4 (EM)	0.271	0.256	+6.1%

Computational Performance

Metric	FieldMem	Traditional
Memory efficiency	42x reduction	baseline
JAX JIT speedup	518x	baseline
Multi-agent sharing	Native	Manual

Scientific Background

FieldMem builds on the Field-Theoretic Context System (FTCS) published in Technical Disclosure Commons. The full methodology and benchmark analysis are available in our arXiv paper.

Links

• Paper: arXiv:cs.ET (FieldMem: Field-Theoretic Memory for AI Agents)
• Prior Work: FTCS Technical Disclosure
• Website: https://rotalabs.ai
• GitHub: https://github.com/rotalabs/rotalabs-fieldmem
• Documentation: https://rotalabs.github.io/rotalabs-fieldmem/
• Contact: research@rotalabs.ai

License

Apache-2.0 License - see LICENSE for details.