BanditDB

BanditDB is an in-memory decision database for AI agents — real-time learning from outcomes. Use it to auto-tune notification timing, model routing, or prompt selection.

BanditDB Skill

BanditDB is a self-hosted decision database. It learns which choice works best for which context through contextual multi-armed bandits — no ML pipeline required.

Setup

Install BanditDB from GitHub releases or run the Docker image (see references/api.md for details). Default port: 8080. Verify by requesting GET /campaigns.

Core Workflow

Three-step loop — create once, then predict and reward repeatedly:

1. Create a campaign — define a campaign ID, the arms (choices), and context feature dimension.
2. Get a prediction — pass a context vector, receive the recommended arm and an interaction ID.
3. Record a reward — report the outcome (0.0–1.0) for the interaction ID.

For full API details, request/response examples, and MCP tool registration, see references/api.md.

Designing Context Vectors

The context vector is the most important design decision. Each float encodes something about the current situation. Normalize values to roughly 0–1 range.

Examples:

• Notification timing: [hour_of_day/24, day_of_week/7, messages_today/10, last_response_delay_mins/60]
• Tool selection: [query_length/500, has_code_mention, has_url, specificity_score]
• Prompt strategy: [task_complexity, domain_familiarity, output_length_needed, structured_output]

Use Cases for OpenClaw Agents

• Smart notifications — learn when/how to reach the user (arms: morning/afternoon/evening, channel variants)
• Tool routing — which tool to use for a query type (arms: web_search/memory/file_lookup/ask)
• Model selection — which model for which task (arms: opus/sonnet/haiku)
• Response style — learn user preferences (arms: brief/detailed/bullet_points)
• Heartbeat frequency — when to check in vs stay quiet

Key Details

• Algorithms: LinUCB (default, supports causal analysis) or Thompson Sampling
• Cold start: meaningful lift typically after 300–1500 interactions depending on noise
• Parquet export available for offline causal analysis (LinUCB only)
• WAL ensures crash recovery — no data loss on restart
• ~10K predictions/s on a single node