AgentRecall-MCP provides persistent, compounding memory for AI agents, enabling them to learn from user corrections and improve over time. It addresses the common issue of AI agents forgetting context, repeating mistakes, and requiring constant re-explanation, thereby reducing wasted tokens and rework. The system is designed for developers, researchers, and power users building or deploying AI agents, offering a significant benefit in agent efficiency and user alignment.

How It Works

AgentRecall operates on the "Intelligent Distance Protocol," aiming to navigate the inherent gap between human thought and AI action. It functions as a learning loop where memory is the mechanism and understanding is the goal. Key to its approach is a five-layer memory pyramid (Journal, Episodic, Palace, Awareness, Insight Index) and a compounding awareness system capped at 200 lines, forcing compression and increasing the value of each stored insight. Corrections are automatically captured, weighted, and recalled before an agent repeats a mistake, allowing the agent to learn the user's specific priorities, communication style, and non-negotiables over sessions. This contrasts with simple logging by focusing on learning from feedback to improve future interactions.

Quick Start & Requirements

• MCP Server Installation: For platforms like Claude Code, Cursor, VS Code, Windsurf, and Codex, use commands like claude mcp add --scope user agent-recall -- npx -y agent-recall-mcp.
• SDK Installation: For JS/TS applications (LangChain, CrewAI, etc.), install via npm: npm install agent-recall-sdk.
• CLI Installation: Install globally via npm: npm install -g agent-recall-cli or use npx agent-recall-cli.
• Prerequisites: Node.js and npm are required for SDK and CLI usage. Specific AI platforms are needed for MCP integration. All data is stored locally in Markdown files, requiring no cloud services or API keys.
• Documentation: Links to Command Reference, Intelligent Distance Protocol, and SDK Design are available.

Highlighted Details

• Persistent, compounding memory with automatic correction capture and recall.
• "Intelligent Distance Protocol" enables agents to learn user's thinking patterns.
• Five-layer memory pyramid and a 200-line awareness cap for knowledge compression.
• Cross-project insight recall allows lessons learned in one project to apply to others.
• Local, zero-cloud, zero-telemetry storage in Markdown format.
• Universal compatibility via MCP, SDK, and CLI across various AI platforms and frameworks.
• Benchmarked significant token savings in complex, multi-session, or multi-agent scenarios.

Maintenance & Community

The project is built by tongwu at Novada. Feedback and issues can be directed via GitHub Issues or email (tong.wu@novada.com).

Licensing & Compatibility

AgentRecall is released under the MIT License, permitting commercial use and integration into closed-source projects without significant restrictions. It is designed for broad compatibility with numerous AI agent frameworks and platforms.

Limitations & Caveats

For very simple, single-session tasks, AgentRecall may introduce overhead. Its compounding benefits are most realized over multiple sessions and with user corrections, meaning initial gains might be less pronounced for ephemeral tasks. Integration requires specific platform support for MCP or a Node.js environment for SDK/CLI usage.