cortex-mem: Give AI Agents Persistent Long-Term Memory

Overview

Cortex Memory helps AI applications retain useful information across sessions instead of relying only on short chat histories. It stores memories in a virtual filesystem and pairs that durable content with vector search for retrieval.

The project is aimed at developers building agents or personalized AI assistants who need memory extraction, organization, and retrieval in one system. Its layered memory model supports retrieving concise summaries before loading full details. See the repository for full documentation and examples.

Key Features

• Organizes memory into three levels: abstracts, overviews, and detailed content.
• Stores memory as Markdown through a cortex:// virtual filesystem.
• Extracts structured memories from conversations using an LLM.
• Supports semantic search with Qdrant and metadata filtering.
• Provides CLI, REST API, MCP server, and Rust integration options.
• Includes tenant-aware memory spaces and a Svelte dashboard for browsing and monitoring.
• Automates indexing and memory updates, with configurable cleanup of weaker memories.

Use Cases

• Agent developers can preserve user preferences and conversation context between sessions.
• Teams building AI assistants can provide personalized retrieval across separate users or agents.
• OpenClaw users can connect an external memory store through the project’s integration ecosystem.
• Developers integrating memory into applications can use the REST API, MCP server, or Rust libraries.

Getting Started

The README lists Rust 1.86 or later, Qdrant 1.7 or later, an OpenAI-compatible LLM endpoint, and an embedding endpoint as prerequisites. Install the CLI with:

cargo install --path cortex-mem-cli

See the README for service and MCP installation, configuration, and usage instructions.

Alternatives

• engram: Engram focuses on retaining project knowledge for AI coding agents in SQLite, while cortex-mem provides file-based memory storage and semantic search for agents and applications.

Considerations

• Running the full system requires Qdrant plus compatible LLM and embedding services, so it is not a standalone memory library with no external dependencies.
• Memory extraction and embeddings depend on configured model endpoints; plan for their availability, cost, and data-handling implications.
• The repository describes the system as production-ready, but the supplied project facts do not establish independent production adoption or validate its benchmark claims.
• The project is actively maintained according to the supplied repository data and is not archived.