Mnemo Cortex: Open-Source AI Agent Memory Coprocessor

What changed

Mnemo Cortex has been released as an open-source project, functioning as a memory coprocessor for AI agents. It is designed to provide persistent recall, semantic search, and crash-safe memory capture for AI systems. A key design principle is that it requires no external hooks to operate. The project emphasizes local execution, with no cloud dependency, telemetry, or investor funding, being developed by a single individual. The latest release is version 2.12.0. The system has been in production use for a working five-agent fleet on two machines since March 2026, ingesting, compacting, and recalling approximately 10,000 memories, and managing around 7,000 verified facts with audit history. It has undergone 84 shipped versions in five months. Features include overnight dreaming for cross-agent synthesis, a document discovery system using an SQLite FTS5 index, an agent-to-agent messaging system called Disco-Bus, and a secure method for transferring memories via USB called Cortex Stick. It also includes a Developer's Passport for safe behavioral-claim ingestion and a Memory Ledger for tamper-evident storage.

Why it matters for builders

AI agents often suffer from amnesia, and Mnemo Cortex aims to be the cure by providing active, persistent memory. Builders can integrate this coprocessor to give their agents the ability to remember past interactions, learned facts, and context across sessions and even across different agents. This is particularly valuable for creating more sophisticated and autonomous AI systems that can maintain state and learn over time without constant re-initialization or loss of information. The project's commitment to being free and locally runnable lowers the barrier to entry for implementing advanced memory capabilities.

Practical impact

Developers can integrate Mnemo Cortex into various AI agent frameworks. The project offers multiple integration paths, including a 60-second install for Claude Code, a drag-and-drop .mcpb bundle for Claude Desktop, and native MCP integration for OpenClaw. It also supports LM Studio, AnythingLLM, Agent Zero, Hermes Agent, Ollama Desktop, and can be used with any local LLM via MCP setup. For ChatGPT, it offers integration through a Custom GPT Actions gate, providing memory without exposing the Mnemo server directly to the internet. The project also provides a "Lane Protocol" document, which serves as an operating practice for running agents with persistent memory, detailing workflow patterns and best practices.

Caveats and source limits

The source material indicates that Mnemo Cortex is a personal project funded by donations, with no venture capital. While it claims production use in a multi-agent fleet, independent benchmarks or extensive third-party validation of its performance and scalability are not provided in the excerpt. The pricing model is donation-based, and specific costs for running the system are not detailed beyond the implication of being free to use. The integration with ChatGPT via a gate has specific OpenAI plan disclaimers that users should review.