bulletproof-memory

What This Skill Does

Bulletproof Memory is a robust architecture for AI agents designed to solve the inherent volatility of LLM context windows. By implementing the Write-Ahead Log (WAL) protocol, this skill moves session-critical information out of the transient, error-prone chat buffer and into a persistent SESSION-STATE.md file. It effectively treats this file as 'Hot RAM,' ensuring that essential facts, decisions, and task parameters survive context window compaction, session restarts, and the natural drift of long-running conversations. By forcing the agent to perform write operations triggered by user input rather than internal memory, the skill ensures that information is captured immediately and reliably.

Installation

To integrate this skill, use the ClawHub CLI: clawhub install openclaw/skills/skills/halthelobster/bulletproof-memory

Once installed, add the required protocol instructions to your AGENTS.md file. Ensure your agent has appropriate filesystem permissions to read and write in the workspace root, as the mechanism relies on real-time modification of the SESSION-STATE.md file.

Use Cases

• Long-term Project Management: Maintaining complex requirements for software development that would otherwise be lost in a 50k+ token conversation.
• Dynamic Correction Handling: Ensuring that user corrections (e.g., 'no, that's the wrong API version') are instantly persisted so the agent never repeats the same mistake.
• Multi-Session Workflows: Resuming complex tasks exactly where you left off after an agent restart or a crash.
• Research Aggregation: Storing discovered facts or URLs in a structured document rather than relying on the agent's 'internal' representation of facts.

Example Prompts

1. "We've decided to pivot the database architecture to PostgreSQL; update the SESSION-STATE.md to reflect this."
2. "Actually, my name is Alex and I prefer tabs over spaces, make sure that's noted for the rest of this session."
3. "Summarize the current status of the project based on the SESSION-STATE file."

Tips & Limitations

• Proactive Management: You can manually edit the SESSION-STATE.md file yourself. The agent will read your changes, reinforcing the shared source of truth.
• Compaction Warning: While this skill protects against compaction-induced forgetting, your agent must be explicitly instructed to prioritize the file's contents over its own fuzzy training memory.
• File Maintenance: Periodically clear or archive old entries in SESSION-STATE.md to prevent the file from becoming too large, which could eventually hit the agent's token limit if it reads the entire file on every turn.