MemBlocks
Modular memory management system for LLMs with isolated, shareable context spaces per conversation.
- LLM Memory
- RAG
- Agentic AI
Overview
MemBlocks is an open-source, pip-installable memory management system for large language models built by Ashish Pandey. It gives every conversation isolated, shareable context spaces so retrieval stays focused, and layers Core, Semantic, Episodic, and Resource memory on top of a Qdrant vector store with an LLM-driven query extraction pipeline and a custom reranker.
What Ashish Pandey built
- Architected a modular memory block system eliminating cross-context retrieval noise
- Designed a four-layer memory model: Core, Semantic, Episodic, and Resources backed by Qdrant
- Built an LLM-powered query extraction pipeline decomposing natural language into intent, entities, and temporal signals
- Implemented a custom reranker weighing semantic relevance, recency, and source credibility
Tech stack
- Python
- LangChain
- Qdrant
- FastAPI
More projects by Ashish Pandey
- TimelyMeet AI-powered meeting management app with multi-agent scheduling and email automation.
- Desktop AI Assistant Voice-controlled system automation agent enabling hands-free control, email, and browser automation.
- Sanket Assistive app for the deaf and hard of hearing with ML-powered environmental sound detection.