ngosangns Knowledge Base

Neural Memory

#technology#memory#code-intelligence

Updated: 2026-09-24

Source

Definition

Neural Memory is an open-source Python library that provides persistent memory for AI agents. It stores experiences as interconnected neurons and recalls them through spreading activation, mimicking how the human brain recalls memories. The CLI stores with nmem remember, recalls with a depth flag, and scores brain health from A to F. nmem sync is the cross-device command, and nmem serve starts the web dashboard. The graph section lists 24 relationship types.

Installation

pip install neural-memory
nmem init --full

Core Tools

Tool Function
nmem_remember Store memory with auto-detected type, tags, connections
nmem_recall Recall via spreading activation
nmem_health Brain health score (A-F) with fix suggestions

CLI Commands

# Store memories
nmem remember "Fixed auth bug with null check in login.py:42"
nmem remember "We decided to use PostgreSQL" --type decision

# Recall through spreading activation
nmem recall "auth bug"
nmem recall "database decision" --depth 2

# Brain health check
nmem brain health

# Sync across devices
nmem sync --full

# Web dashboard
nmem serve

Python API

from neural_memory import Brain
from neural_memory.storage import InMemoryStorage
from neural_memory.engine.encoder import MemoryEncoder
from neural_memory.engine.retrieval import ReflexPipeline

# Initialize brain
brain = Brain(storage=InMemoryStorage())

# Store memory
brain.remember(
    content="Fixed auth bug with null check",
    memory_type="bugfix",
    tags=["auth", "bug", "login"],
    connections=["authentication", "validation"]
)

# Recall with spreading activation
results = brain.recall(
    query="auth bug",
    depth=2,  # Multi-hop
    threshold=0.5
)

# Check brain health
health = brain.health()
print(health.score)  # A-F scale

Configuration

# ~/.neuralmemory/config.toml
storage_backend = "infinitydb"

Cloud Sync

Deploy to your own Cloudflare account (free tier) - data is never stored by the service.

nmem sync --full

Graph-Based Architecture

Neural Memory uses a graph-based approach with 24 explicit relationship types:

CAUSED_BY, LEADS_TO, RELATED_TO, SIMILAR_TO,
DEPENDS_ON, IMPLEMENTS, EXTENDS, OVERRIDES,
CALLS, CALLED_BY, IMPORTED_BY, DEFINES,
USES, USED_BY, CONTAINS, CONTAINED_IN,
FOLLOWS, PRECEDES, ANSWERS, ASKS,
SUPPORTS, OPPOSES, ENABLES, BLOCKS

Spreading Activation Algorithm

1. Query → Find matching neurons (activation sources)
2. Spread → Propagate activation through connections
3. Decay → Activation decreases with distance
4. Reinforce → Strong connections strengthen
5. Consolidation → Memories merge over time

Memory Compression Tiers

Tier Description
Full Complete memory with all details
Summary Condensed version
Essence Core facts only
Ghost Metadata only
Reference Pointer to external storage

Setup by Tool

Claude Code

/plugin marketplace add nhadaututtheky/neural-memory

MCP Clients

pip install neural-memory
# Add to your MCP config

OpenClaw

pip install neural-memory && npm install -g neuralmemory

56 MCP Tools

Neural Memory provides 56 MCP tools for:

  • Memory operations
  • Brain management
  • Sync operations
  • Health checks

Tech Stack

  • Python 3.11+
  • SQLite / InfinityDB
  • FastAPI
  • React dashboard
  • MCP protocol
  • MIT License

Strengths

Strength Description
Graph-based True graph memory, not a vector store
Fully offline No dependency on external APIs
Good free tier SQLite + keyword search at no cost
Multi-hop reasoning Outstanding associative recall
Privacy-focused User data never stored

Weaknesses

Weakness Description
Pro requires payment InfinityDB costs money for semantic search
Learning curve Requires understanding graph-based memory
Limited integrations Fewer than some alternatives

When to Use

  • Offline AI agents: You don't want to depend on cloud APIs
  • Cost-sensitive projects: You don't want to pay for a vector database
  • Graph memory needs: Multi-hop reasoning is essential
  • Privacy-first applications: User data must not be stored
  • Persistent context: You need context across sessions

References:

See also: Mem0 · Memory Tool Comparison · Graph Comparison