Summary

memX is a self-learning, self-maintaining memory plugin designed to enhance AI agents by converting completed work into structured, searchable memory. It intelligently injects only the necessary evidence for an agent's current query, improving efficiency and context management. The project targets developers building AI agents, offering native integration with popular frameworks like Codex, Claude Code, and OpenClaw, while also supporting any MCP-compatible client.

How It Works

memX operates by capturing an agent's completed tasks and outputs, transforming them into a persistent, structured, and searchable memory layer. It employs native hooks for seamless integration with agents like Codex and Claude Code, automating the process of memory recall and turn capture. The core innovation lies in its ability to dynamically filter and inject only the most relevant pieces of evidence for the agent's immediate needs, preventing context window bloat and improving decision-making. The system is designed for flexibility, supporting various LLM providers and local embedding models.

Quick Start & Requirements

Installation and setup are primarily managed via npx, pulling directly from GitHub. The primary command is npx -y -p github:NeoLi00/memX memx quickstart <agent-type>, with specific types including claude-code, codex, openclaw, and mcp. Prerequisites include Node.js 22.14+ or Node 24. Python 3 is required for the default local embedding runtime. OpenClaw integrations necessitate OpenClaw 2026.3.25+. Configuration requires specifying LLM provider details (--llm-provider, --llm-base-url, --llm-model, --llm-api-key or environment variable). The default embedding uses sentence-transformers-local with intfloat/multilingual-e5-small.

Highlighted Details

• Achieves a 94.2% R@3 success rate on the LongMemEval-S benchmark for long-context memory retrieval.
• Demonstrated 100% success across 30 real-world engineering cases, each involving over 20 turns.
• Provides native lifecycle hooks for Codex and Claude Code, enabling automatic memory recall and turn capture.
• Offers broad compatibility with any MCP-compatible client through its local memory layer.

Maintenance & Community

No specific details regarding maintainers, community channels (like Discord/Slack), or roadmap were present in the provided README snippet.

Licensing & Compatibility

The provided README snippet does not specify a software license. Consequently, compatibility for commercial use or closed-source linking cannot be determined without further information.

Limitations & Caveats

By default, native memories are host-scoped, meaning agents like Codex and Claude Code maintain separate local databases unless explicitly configured otherwise. The quickstart process will halt if the default memX service port (3787) is already in use by an unmanaged service, requiring manual intervention or port reassignment. Generic MCP quickstarts default to --mcp-tools full, potentially exposing a broader toolset than native integrations.