mumpu

Middleware for Universal Memory Persistence and Understanding

A transparent HTTP relay proxy that adds long-term memory to any LLM application. Point your tools at mumpu and it remembers everything across sessions — extracts knowledge, builds connections, and injects relevant context automatically.

  Your tools ──► mumpu (:8420) ──► OpenAI / Anthropic / Gemini
                   │
              Extracts memories
              Builds connections
              Injects context

Quick Start

pip install -e ".[dev]"

# Start the proxy with TUI dashboard
mumpu start

# In another terminal, use Claude through the proxy
export ANTHROPIC_BASE_URL=http://localhost:8420
mumpu claude

Open http://localhost:8420/dashboard to see the memory graph grow in real time.

How It Works

• Middleware — relay between your tools and the API
• Universal — works with any tool, any provider (OpenAI, Anthropic, Gemini)
• Memory — extracts and stores knowledge from conversations
• Persistence — memories survive across sessions in SQLite
• Understanding — smart retrieval with graph-based connections, not dumb storage

Documentation

• Architecture — system design, data flow, module map
• Setup & Configuration — installation, environment variables, CLI reference
• Agent Handoff — current state, known issues, patterns for contributing
• Paper Gaps — what the research papers propose that we haven't built yet
• Migration Plan — SQLite migration details (completed)