Skip to content

Auto-Capture Rules

The real power of Open Brain is that agents use it without being asked. This is achieved through behavioral rules injected into each agent's system prompt or rules file.

The Philosophy

The user's only job is to work. The brain's job is to remember everything.

Agents should:

  • Never announce memory operations. Capture and recall silently.
  • Never ask "should I remember this?" Just do it.
  • Always pass source on capture calls so you can see where memories came from

When Agents Auto-Capture

Agents call capture_context automatically:

  • After completing any task or feature
  • When a technical decision is made (what was chosen, what was rejected, why)
  • When a bug is diagnosed and fixed (root cause, symptoms, the fix)
  • When something about user preferences or the project is learned
  • At natural conversation checkpoints or end of long sessions

When Agents Auto-Recall

Agents call search automatically:

  • At the start of any task to check for prior context
  • When the user mentions a person, project, or technology
  • Before making a recommendation, to check if a past decision already exists

Setting Up Rules

Each AI client has a different mechanism for persistent behavioral instructions:

Windsurf

Copy prompts/windsurf-rules.md into .windsurfrules at the root of your project, or into Windsurf's global rules at:

C:\Users\<USERNAME>\.codeium\windsurf\memories\global_rules.md

Cursor

Copy prompts/cursor-rules.md into:

  • .cursorrules at the root of your project, or
  • Settings (Ctrl+Shift+J) > General > Rules for AI (global)

Claude Desktop

Copy prompts/claude-desktop.md into:

Settings > Custom Instructions

Claude Code

Add to CLAUDE.md at the root of your project. See prompts/generic-system-prompt.md for the template.

What Gets Captured

A typical capture_context call from an agent might look like:

Session context: Fixed the authentication bug in login.py.
Root cause was a missing await on the token validation call.
The function returned a coroutine object instead of the actual token.
Decided to add an integration test that hits the real auth endpoint
rather than mocking it, since we got burned by mock divergence before.

Open Brain decomposes this into atomic memories:

  1. Bug fix: Missing await on token validation in login.py (type: insight)
  2. Decision: Use integration tests for auth, not mocks (type: decision)

Each gets its own embedding, metadata, and dedup check.