Most AI agents today appear to “remember” things, but the reality is that LLMs themselves are completely stateless. The memory illusion usually comes from external systems:
• conversation history stored in a DB
• vector retrieval (RAG)
• summarization pipelines
• fact extraction services
These layers are often glued together with frameworks and APIs.
I built Mumpix to simplify that stack.
Mumpix is a lightweight memory engine designed specifically for AI agents. It runs on both the frontend and backend, and the goal is to provide a simple persistent memory layer without requiring vector databases, servers, or complex orchestration.
The core ideas behind the project:
• Structured agent memory using hierarchical keys
• Persistent state across sessions (browser or Node)
• Deterministic reads/writes instead of probabilistic vector search
• Portable memory snapshots that can be exported or replayed
• No infrastructure required to get started
It’s designed to behave more like SQLite for AI memory than a typical AI platform.
Some things it enables:
• agents that remember user preferences locally
• deterministic state tracking for agent workflows
• offline AI apps with persistent memory
• explainable responses (tracking which keys were read)
The core engine is intentionally small and dependency-free so it can run anywhere.
As of v1.17, Mumpix works across the full stack:
• Browser (IndexedDB persistence)
• Node.js
• optional sync layers
I’d love feedback from people building agents or local AI systems.