MindCache offers a structured long-term memory engine specifically designed for production-grade large language model (LLM) agents. Unlike conventional vector search databases or basic summary storage systems, MindCache features a self-restructuring hierarchical memory ontology that captures explicit decision tracking, thereby outperforming traditional flat retrieval systems demonstrated on the BEAM QA benchmark.

Key Features

Advanced Memory Architecture

MindCache redefines long-term memory by organizing it into a structured, self-organizing architecture optimized for intelligent reasoning. This sophisticated structure is composed of the following functionalities:

• Self-Restructuring Hierarchical Ontology: Continuously adapts and organizes memories, ensuring relevant information is easily accessible while minimizing redundancy.
• Four Distinct Memory Types: Memory is stored as distinct categories, which include:
  • 📝 Episodic: Captures interactive events and session context.
  • 🧠 Knowledge: Holds factual statements and concepts relevant to the domain.
  • 👤 User: Includes identity facts and behavioral attributes.
  • ⚖️ Decision: Keeps track of explicit choices and directives made by the user.

Intelligent Ingestion and Retrieval

• Smart Ingestion Mechanics: Raw conversations are embedded, processed in batches with a focus on efficient routing, ensuring contextually relevant information is captured effectively.
• Dynamic Retrieval Pipeline: Queries are processed through a multi-stage engine that utilizes optimal classification and direct boosting based on user decisions, resulting in precision retrieval.

Usage Example

To use MindCache, the integration involves initial setup through installation and simple code snippets as demonstrated below:

from mindcache import MindCache

# Initialize the MindCache client with the SQLite database backend
mc = MindCache(db_path="my_memory.db", provider="gemini", model_name="gemini-2.5-flash")

# Add a conversation turn
job_id = mc.add([
    {"role": "user", "content": "I prefer Python and FastAPI for backend development."},
    {"role": "assistant", "content": "Got it! I will remember your preferences."}
], user_id="alice")

# Process the queue to update the memory
mc.process_queue(user_id="alice")

# Retrieve relevant context for future prompts
context = mc.search("What is Alice's preferred database?", user_id="alice")
print(context)

Performance Evaluation

MindCache has been rigorously benchmarked against the BEAM quality assessment, achieving significantly improved accuracy metrics while maintaining a reduced retrieval latency. For instance:

• Average BEAM-1M Accuracy: 87.0%, surpassing the previous benchmark of 64.1%.

This efficiency ensures MindCache stands out as an optimal solution for maintaining long-term memory in AI systems, focusing on performance and user-context relevancy.

For a more detailed exploration of MindCache's architectural design, see the blog post: Building MindCache: Designing an Agentic Memory System for Long-Term AI.