ICM provides AI agents with persistent, structured memory, moving beyond transient context windows. It addresses the need for agents to retain and recall information effectively, enhancing their capabilities for complex tasks. The target audience includes AI developers and power users building sophisticated AI agents, offering benefits such as improved efficiency, knowledge retention, and multi-agent collaboration.

How It Works

ICM employs a dual-model architecture for AI agent memory: 'Memories' capture episodic, temporally decaying information, with 'critical' memories persisting indefinitely. 'Memoirs' build permanent knowledge graphs using concepts and typed relations. This is underpinned by SQLite with FTS5 and sqlite-vec, enabling a hybrid search combining BM25 (30%) and cosine similarity (70%). The system prioritizes local-first operation, a single binary, and zero external dependencies for its core functionality.

Quick Start & Requirements

Installation is streamlined via Homebrew (brew install icm), a curl/PowerShell script, or from source (cargo install --path crates/icm-cli). The icm init command automatically configures integration with 17 supported AI agent tools (MCP-compatible). While the core binary has zero dependencies, full functionality relies on having the target AI agent tools configured. Embedding models can be disabled for offline use.

Highlighted Details

• Dual Memory Architecture: Episodic (decaying) and Semantic (knowledge graph) models.
• Hybrid Search: Weighted combination of BM25 and cosine similarity within SQLite.
• Zero-LLM Auto-Extraction: Three layers capture memories from tool output, transcripts, and prompts without LLM calls.
• Unified Multi-Agent Memory: A single SQLite database allows seamless memory sharing across 17 integrated AI tools.
• Performance Benchmarks: Demonstrates high storage/search speeds and significant gains in agent efficiency and knowledge retention across sessions.

Maintenance & Community

The project is maintained by rtk-ai. No specific community channels (e.g., Discord, Slack) or prominent contributor details are listed in the provided README.

Licensing & Compatibility

ICM is licensed under the Apache-2.0 license, which is permissive for commercial use and integration into closed-source projects. Its local-first, zero-dependency design facilitates broad compatibility.

Limitations & Caveats

Full integration with AI agents requires configuring specific MCP-compatible tools. Optional embedding models must be downloaded or explicitly disabled. While the core is zero-dependency, leveraging the full suite of features necessitates setup across various agent environments.