Swappable Brains for AI Agents
CA: F8nY58QdpSgkwWRiKezBLdneVGQt5HwHMSUH5q3ApumpBrain Cartridges for AI Agents
Membot is an MCP server that gives AI agents portable, hot-swappable memory through Brain Cartridges. Unlike standard AI memory that forgets between sessions, Membot stores knowledge on a neuromorphic substrate — a 16-million neuron lattice with real physics. Agents can swap cartridges to gain deep domain knowledge instantly, from medical literature to legal documents, without retraining or losing context. One server, many agents, many domains.
Claude Code OpenClaw Base / $MEM
Most AI memory relies on a single math operation — cosine similarity — to find related content. Membot blends three independent signals: semantic embeddings, binary population codes derived from the neural structure of each memory, and keyword matching. The result is search that catches what any single method would miss.
Behind the search layer sits a neuromorphic lattice — a hybrid Hopfield network with Hebbian learning, energy dynamics, and content-addressable recall. This isn't a database. It's a substrate that forms genuine associations between stored patterns. Present a partial cue and the physics converges to the right memory, even with noise or missing information. Your AI doesn't just find similar words — it finds contextual connections.
Technical Explanation and demos
Built on the Vector+ LatticeRunner Engine, Membot is designed for the next generation of AI workflows. It operates locally or in the cloud as an MCP server, natively compatible with Claude Code, OpenClaw, and any MCP-compatible agent framework.
Three-Signal Search
Blends embedding cosine similarity (70%), binary Hamming distance on neuromorphic population codes (30%), and keyword reranking to surface results that any single method would miss. No GPU required.
Swappable Cartridges
Memory is stored in self-contained Brain Cartridges (.npz files with SHA256 integrity verification). Mount a cartridge, search it, swap to another — each agent gets its own isolated session state.
Neuromorphic Recall
A 64x64 grid of 64 regions (16 million neurons) with Hebbian weights provides content-addressable memory. Present a noisy or partial cue and the attractor dynamics converge to the correct stored pattern. Validated to 1 million patterns.
Flexible Transport
Supports stdio for local agent subprocesses and HTTP/SSE for remote deployments. Run it on a laptop or a cloud server — same codebase, same cartridges.
Production-Ready
Built-in rate limiting, API key authentication, session-based state isolation for up to 0 concurrent agents, and a live operational dashboard showing cartridge status and agent activity in real time.