neural-memory-enhanced（神经记忆增强）

NeuralMemory — Associative Memory for AI Agents

A biologically-inspired memory system that uses spreading activation instead of keyword/vector search. Memories form a neural graph where neurons connect via 20 typed synapses. Frequently co-accessed memories strengthen their connections (Hebbian learning). Stale memories decay naturally. Contradictions are auto-detected.

Why not just vector search? Vector search finds documents similar to your query. NeuralMemory finds *conceptually related* memories through graph traversal — even when there's no keyword or embedding overlap. "What decision did we make about auth?" activates time + entity + concept neurons simultaneously and finds the intersection.

Setup

1. Install NeuralMemory

pip install neural-memory
nmem init

This creates ~/.neuralmemory/ with a default brain and configures MCP automatically.

2. Configure MCP for OpenClaw

Add to your OpenClaw MCP configuration (~/.openclaw/mcp.json or project openclaw.json):

{
  "mcpServers": {
    "neural-memory": {
      "command": "python3",
      "args": ["-m", "neural_memory.mcp"],
      "env": {
        "NEURALMEMORY_BRAIN": "default"
      }
    }
  }
}

3. Verify

nmem stats

You should see brain statistics (neurons, synapses, fibers).

Tools Reference

Core Memory Tools

Intelligence Tools

Management Tools

Workflow

At Session Start

1. Call nmem_context to inject recent memories into your awareness
2. If user mentions a specific topic, call nmem_recall with that topic

During Conversation

1. When a decision is made: nmem_remember with type="decision"
2. When an error occurs: nmem_remember with type="error"
3. When user states a preference: nmem_remember with type="preference"
4. When asked about past events: nmem_recall with appropriate depth

At Session End

1. Call nmem_auto with action="process" on important conversation segments
2. This auto-extracts facts, decisions, errors, and TODOs

Examples

Remember a decision

nmem_remember(
  content="Use PostgreSQL for production, SQLite for development",
  type="decision",
  tags=["database", "infrastructure"],
  priority=8
)

Recall with spreading activation

nmem_recall(
  query="database configuration for production",
  depth=1,
  max_tokens=500
)

Returns memories found via graph traversal, not keyword matching. Related memories (e.g., "deploy uses Docker with pg_dump backups") surface even without shared keywords.

Trace causal chains

nmem_recall(
  query="why did the deployment fail last week?",
  depth=2
)

Follows CAUSED_BY and LEADS_TO synapses to trace cause-and-effect chains.

Auto-capture from conversation

nmem_auto(
  action="process",
  text="We decided to switch from REST to GraphQL because the frontend needs flexible queries. The migration will take 2 sprints. TODO: update API docs."
)

Automatically extracts: 1 decision, 1 fact, 1 TODO.

Key Features

• Zero LLM dependency — Pure algorithmic: regex, graph traversal, Hebbian learning
• Spreading activation — Associative recall through neural graph, not keyword/vector search
• 20 synapse types — Temporal (BEFORE/AFTER), causal (CAUSED_BY/LEADS_TO), semantic (IS_A/HAS_PROPERTY), emotional (FELT/EVOKES), conflict (CONTRADICTS)
• Memory lifecycle — Short-term → Working → Episodic → Semantic with Ebbinghaus decay
• Contradiction detection — Auto-detects conflicting memories, deprioritizes outdated ones
• Hebbian learning — "Neurons that fire together wire together" — memory improves with use
• Temporal reasoning — Causal chain traversal, event sequences, temporal range queries
• Brain versioning — Snapshot, rollback, diff brain state
• Brain transplant — Transfer filtered knowledge between brains
• Vietnamese + English — Full bilingual support for extraction and sentiment

Depth Levels

Notes

• Memories are stored locally in SQLite at ~/.neuralmemory/brains/<brain>.db
• No data is sent to external services (unless optional embedding provider is configured)
• Brain isolation: each brain is independent, no cross-contamination
• nmem_remember returns fiber_id for reference tracking
• Priority scale: 0 (trivial) to 10 (critical), default 5
• Memory types: fact, decision, preference, todo, insight, context, instruction, error, workflow, reference