CaviraOSS/LongMemory

What it solves

OpenMemory provides a cognitive memory engine for LLMs and AI agents to overcome the "goldfish" effect—where models forget everything between messages. Unlike standard RAG (Retrieval-Augmented Generation) or simple vector databases, it implements a structured memory system that understands the nature of information (facts, events, preferences) and how it evolves over time.

How it works

The system uses Hierarchical Memory Decomposition to categorize information into five sectors: episodic (events), semantic (facts), procedural (skills), emotional (feelings), and reflective (insights). It combines vector search with a temporal knowledge graph to track when facts were true and how they change. A composite scoring system—using salience, recency, and coactivation—determines what is recalled, while a decay engine manages adaptive forgetting.

It can be deployed as a local-first SDK (Python or Node.js) using SQLite/Postgres, or as a centralized backend server with an HTTP API and an MCP (Model Context Protocol) server for integration with IDEs like Cursor and Windsurf.

Who it’s for

Developers building AI agents, copilots, personalized journaling systems, and coding assistants who need their applications to have durable, explainable, and self-hosted long-term memory.

Highlights

• Multi-sector memory: Categorizes data into episodic, semantic, procedural, emotional, and reflective sectors.
• Temporal Reasoning: Uses a knowledge graph to track truth windows (valid_from/valid_to) and reconstruct timelines.
• Explainable Recall: Provides "waypoint" traces to show exactly why a specific memory was recalled.
• Local-first & Self-hosted: Supports local SQLite by default, avoiding vendor lock-in and privacy concerns.
• Broad Integrations: Native support for LangChain, CrewAI, AutoGen, and MCP, with connectors for GitHub, Notion, and Google Drive.