NeuroVault

OpenClaw plugin for BrainBox. Gives your agent memory that learns — combines Hebbian learning (procedural memory) with VaultGraph (knowledge graph).

Quick Start

# 1. Clone and install
git clone https://github.com/thebasedcapital/neurovault.git
cd neurovault && npm install

# 2. Install backends (both optional — install at least one)
cargo install --git https://github.com/thebasedcapital/vaultgraph  # knowledge graph
npm install -g brainbox-hebbian                                     # Hebbian memory

# 3. Register with OpenClaw
openclaw --dev plugins install -l $(pwd)

Then add to ~/.openclaw-dev/openclaw.json:

{
  "plugins": {
    "slots": {
      "memory": "neurovault"
    }
  }
}

Restart the gateway and you're done:

launchctl kickstart -k gui/$(id -u)/ai.openclaw.dev

Verify

openclaw --dev agent --agent main --local -m "hello"
# Check logs for [neurovault]:
tail -20 ~/.openclaw-dev/logs/gateway.log

What It Does

Before every agent prompt, NeuroVault queries two memory systems and injects relevant context:

[neurovault] Unified memory context for this session:

[vaultgraph] Relevant memory files for this task:
  - trading-polymarket (score: 100%, ~1699tok)
  - general-lessons (score: 79%, ~1072tok)

[brainbox] Neural recall for this task:
  - ~/project/src/market.py (confidence: 82%, myelin: 45%)
  - ~/project/src/redeem.py (confidence: 68%, myelin: 38%)

After every tool call, it learns which files were accessed together (Hebbian learning). Over time, it builds muscle memory for your codebase.

System	Type	What It Learns	Speed
VaultGraph	Declarative (what you know)	Knowledge graph over markdown notes	<5ms
BrainBox	Procedural (how you work)	File access patterns, error-fix pairs, tool chains	~100ms

Configuration

All optional — defaults work out of the box:

Variable	Default	Description
NEUROVAULT_ENABLED	true	Master enable/disable
NEUROVAULT_VAULT_PATH	~/.openclaw/memory	Markdown vault directory
NEUROVAULT_VG_BUDGET	3000	Max tokens for VaultGraph context
NEUROVAULT_BB_BUDGET	5000	Max tokens for BrainBox context
NEUROVAULT_MIN_CONFIDENCE	0.5	Minimum BrainBox confidence to show

Architecture

OpenClaw before_agent_start
    |
    |---> VaultGraph (Rust subprocess, <5ms)
    |     Spreading activation over wikilink graph
    |
    +---> BrainBox (SQLite + Hebbian engine, ~100ms)
          Neural recall over file co-access patterns
    |
    v
Combined context injected as system message

OpenClaw after_tool_call
    |
    +---> BrainBox records file access (Hebbian learning)

OpenClaw agent_end
    |
    +---> Captures facts/preferences as semantic neurons

Hooks & Tools

Hook	Event	Purpose
before_agent_start	Every prompt	Inject relevant context
after_tool_call	Every tool use	Learn file access patterns
agent_end	Session end	Capture conversation highlights

Tool	Description
neurovault_recall	Manually query memory
neurovault_stats	Show memory statistics

Related

• BrainBox — Core Hebbian memory engine (also works standalone with Claude Code, Kilo)
• VaultGraph — Knowledge graph CLI for markdown vaults
• OpenClaw — AI agent platform

License

MIT