verygoodplugins/automem

Long-term memory for AI assistants. Graph + vector store that recalls decisions, relationships, and context across sessions.

What it solves

AutoMem provides long-term memory for AI assistants, addressing the limitations of standard RAG and vector databases. While traditional RAG often returns documents based on keyword similarity without understanding context or learning over time, AutoMem stores both typed relationships and embeddings to recall not just isolated facts, but the reasoning, alternatives, and principles behind decisions.

How it works

AutoMem uses a dual-storage architecture to manage memory:

  • Graph Layer (FalkorDB): Acts as the canonical record, storing memories as nodes connected by 11 specific relationship types (e.g., PREFERS_OVER, LEADS_TO, CONTRADICTS).
  • Vector Layer (Qdrant): Stores embeddings for every memory to enable semantic search.

Recall is performed via a hybrid query that combines semantic similarity, graph traversal, temporal alignment, and importance. The system also implements neuroscience-inspired consolidation cycles (Decay, Creative, Cluster, and Forget) to strengthen important connections and fade irrelevant information over time. It can be integrated into AI clients via a Model Context Protocol (MCP) bridge or a direct REST API.

Who it’s for

  • Solo developers, prosumers, or small teams who want a persistent memory layer across different AI tools (e.g., Claude, Cursor, ChatGPT).
  • Users who prefer self-hosting their data via Docker or Railway to avoid vendor lock-in.
  • Developers building AI agents that require large-context scaling and efficient retrieval of associative memories.

Highlights

  • Hybrid Retrieval: Combines graph-based multi-hop discovery with vector embeddings for deeper context.
  • Typed Relationships: Supports 11 authorable edge types to capture complex logic like causality and contradictions.
  • Biological Consolidation: Uses periodic cycles to mimic human memory, pruning low-relevance data and clustering patterns.
  • High Efficiency: Optimized for large-context scaling, feeding answerers fewer tokens while maintaining high retrieval performance on large datasets.