hebbmem 0.3.0

Hebbian memory for AI agents — memories that fire together wire together.

hebbmem

Hebbian memory for AI agents — memories that fire together wire together.

Install

pip install hebbmem

For higher-quality semantic embeddings:

pip install hebbmem[ml]

Quick Start

from hebbmem import HebbMem

mem = HebbMem()

# Store memories
mem.store("Python is great for data science", importance=0.8)
mem.store("JavaScript runs in the browser", importance=0.5)
mem.store("Neural networks learn from data", importance=0.7)

# Time passes, memories decay
mem.step(5)

# Recall activates related memories through the graph
results = mem.recall("machine learning with Python", top_k=3)
for r in results:
    print(f"{r.content} (score={r.score:.3f})")

How It Works

hebbmem replaces flat vector storage with three neuroscience mechanisms:

Decay — Memories fade over time unless reinforced, following the Ebbinghaus forgetting curve. Recent and frequently accessed memories stay strong.

Hebbian Learning — Memories recalled together strengthen their connections. "Neurons that fire together wire together." Over time, the graph learns which memories are related through usage, not just embedding similarity.

Spreading Activation — Recalling one memory activates related ones through the graph, surfacing connections that keyword or vector search alone would miss.

Persistence

Save and restore memory state across sessions:

mem.save("agent_memory.hebb")

# Later...
mem = HebbMem.load("agent_memory.hebb", encoder="hash")

Uses SQLite internally — single file, zero dependencies, crash-safe.

Configuration

from hebbmem import HebbMem, Config

config = Config(
    activation_decay=0.9,    # how fast activation fades (0-1)
    strength_decay=0.999,    # how fast long-term strength fades (0-1)
    hebbian_lr=0.2,          # learning rate for co-activation (0-1)
    spread_factor=0.5,       # energy spread per hop (0-1)
    max_hops=3,              # BFS depth for spreading activation
    scoring_weights={        # recall ranking formula
        "activation": 0.4,
        "similarity": 0.35,
        "strength": 0.15,
        "importance": 0.1,
    },
)
mem = HebbMem(encoder="hash", config=config)

Batch Store

ids = mem.store_batch(
    ["memory one", "memory two", "memory three"],
    importances=[0.9, 0.5, 0.3],
)

Thread Safety

All public methods are thread-safe (protected by threading.RLock). Safe to use from multiple threads in agent frameworks.

Logging

hebbmem uses stdlib logging. Enable debug output:

import logging
logging.basicConfig(level=logging.DEBUG)

Benchmark

hebbmem outperforms flat vector search (pure cosine similarity) on key scenarios:

Scenario	Metric	hebbmem	Baseline	Delta
Associative	Precision@5	0.56	0.44	+27.3%
Associative	Assoc. Hit Rate	0.13	0.00	—
Noise	Precision@5	0.60	0.56	+7.1%
Contradiction	Precision@1	0.50	0.00	—
Contradiction	MRR	0.75	0.50	+50.0%
Temporal	Precision@5	0.36	0.36	tied

Key wins: contradiction handling (50% MRR improvement — newer memories rank higher) and associative recall (finds indirectly connected memories that baseline misses entirely). Results use HashEncoder; advantages are larger with semantic embeddings.

Run yourself: uv run python benchmarks/run_benchmark.py — see benchmarks/ for methodology.

Examples

See examples/ for runnable scripts:

• basic_usage.py — store, decay, recall, forget
• agent_integration.py — using hebbmem as an agent's memory backend
• custom_config.py — tuning decay, Hebbian learning, and scoring weights

Links

• GitHub
• Changelog
• Codepawl