Open-Source Memory for AI Coding Agents, 'Mnemos', Released

Mnemos is a memory layer that helps AI coding agents retain project conventions, fixes, and key decisions even as sessions change.

While existing memory tools focused on 'retrieving' information, Mnemos focuses on actually getting agents to record information. When an agent fixes a specific error, it saves it as structured data (what was tried, why it failed, the solution, etc.), and when fixes on the same topic recur, it promotes them into a single 'skill' for management.

Key Features and Technical Specifications:

• Lightweight and Standalone: Provided as a single binary of about 15MB written in Go, it runs on Linux, macOS, and Windows. Data is managed via SQLite.
• Minimal LLM Dependency: Rather than running an LLM directly inside the memory layer, it is designed using clustering and template-based methods to ensure consistency of results and enable auditing.
• Compatibility: Supports major AI coding tools such as Claude Code, Claude Desktop, Cursor, Windsurf, and OpenAI Codex CLI.
• Verification Tooling: The mnemos verify command allows quantitative measurement of memory retrieval, changes in agent behavior, and recording capability.

In its own tests, it recorded an overall performance improvement of about 40% in terms of applying project-specific conventions and the agent's self-recording capability.