rtk-ai/icm

What it solves

ICM (Infinite Context Memory) provides a permanent, unified memory layer for AI agents. It eliminates the need to repeatedly explain the same project context, preferences, or technical decisions to different AI tools (e.g., Claude Code, Cursor, Gemini), allowing them to share a single, persistent knowledge base.

How it works

ICM uses a single SQLite database with hybrid search (BM25 and cosine similarity) to store information in three distinct formats:

• Memories: Episodic data (decisions, errors) that uses a temporal decay system. Importance levels (critical, high, medium, low) determine how quickly a memory fades unless it is frequently accessed.
• Memoirs: Permanent knowledge graphs where concepts are linked by typed relations (e.g., depends_on, contradicts).
• Transcripts: Verbatim session replays of agent interactions for audit and review.

It integrates with AI tools via a Model Context Protocol (MCP) server, CLI commands, or system hooks that can automatically extract facts from tool outputs and inject context into prompts.

Who it’s for

Developers and power users who use multiple AI coding assistants or agents and want a consistent, cross-tool memory that persists across sessions and different LLM providers.

Highlights

• Cross-Tool Synchronization: A single database shared across 18+ supported tools including Claude Code, Cursor, and Aider.
• Zero Dependencies: Ships as a single binary using Rust and SQLite.
• Hybrid Search: Combines keyword search (FTS5) with semantic vector search via ONNX Runtime.
• Closed-Loop Learning: Includes a feedback system to record and search past mistakes to prevent recurring errors.
• TUI Dashboard: An interactive terminal interface for managing topics, memories, and knowledge graphs.