Agent Skill Integration

Overview

HEBBS ships a universal SKILL.md file that teaches any AI agent how to use HEBBS as a cognitive memory backend. The skill works with OpenClaw, Cursor, Claude Code, Codex, or any agent that can run shell commands.

The agent uses the hebbs CLI directly — no plugin, SDK, or custom integration needed.

Installation

1. Install HEBBS: brew install hebbs-ai/tap/hebbs (or curl -sSf https://hebbs.ai/install | sh)
2. Initialize a vault: hebbs init . --provider openai --key $OPENAI_API_KEY (--model is optional, defaults per provider)
3. Index your files: hebbs index .
4. Copy hebbs/skills/hebbs/SKILL.md into your agent’s skill directory

For OpenClaw:

cp hebbs/skills/hebbs/SKILL.md ~/.openclaw/skills/hebbs/SKILL.md
openclaw skills list  # verify it appears

How It Works

The skill teaches the agent five core operations:

Operation	Purpose	When to use
remember	Store a memory	User shares a fact, preference, decision
recall	Retrieve relevant memories	Before answering questions
insights	Retrieve consolidated knowledge	Before recall, for distilled answers
forget	Remove outdated memories	User corrects or retracts information
prime	Load entity context	Start of a conversation

HEBBS handles reflection and contradiction detection autonomously using its configured LLM. The agent can optionally trigger reflection manually:

Operation	Purpose	When to use
reflect	Trigger reflection for an entity	Optional one-shot trigger when the agent wants immediate consolidation

Example: OpenClaw Agent Loop

A typical agent conversation flow with HEBBS:

1. User opens chat — Agent runs hebbs prime <entity> to load context
2. User sends message — Agent runs hebbs recall "<message>" --format json to find relevant memories
3. Agent responds using recalled context
4. After response — Agent runs hebbs remember "<key fact from conversation>" --importance 0.7 --entity-id <entity>
5. Before deep questions — Agent runs hebbs insights --entity-id <entity> to check consolidated knowledge

Periodically, HEBBS automatically consolidates memories into insights using the configured LLM (threshold: 20+ memories per entity, or 1 hour interval). The agent can trigger this manually with hebbs reflect --entity-id <id> if immediate consolidation is needed.

Configuration

The agent connects to the HEBBS server at localhost:50051 by default. Override with:

hebbs --endpoint myhost:50051 recall "query"

For authenticated deployments, set the API key:

hebbs --api-key $HEBBS_API_KEY recall "query"

HEBBS requires an LLM provider for autonomous reflection and contradiction detection. Configure during initialization:

hebbs init --provider openai --key $OPENAI_API_KEY

--model is optional (defaults per provider). LLM and embedding config are saved to ~/.hebbs/config.toml and inherited by all vaults.

Or set in ~/.hebbs/config.toml directly:

[llm]
provider = "openai"
model = "gpt-4o-mini"
api_key = "sk-proj-your-key-here"

[embedding]
provider = "openai"
model = "text-embedding-3-small"
api_key = "sk-proj-your-key-here"
dimensions = 1536

Tips

• Always use --format json when the agent parses output
• Set importance based on content type: 0.9 for corrections, 0.8 for preferences, 0.5 for observations, 0.3 for transient context
• Use entity IDs to scope memories per user, project, or domain
• Check insights before raw recall — insights are higher-quality, distilled knowledge