Memory

What it does

A persistent typed memory layer accessed via the open Model Context Protocol. Any MCP-aware agent (Claude Code, Cursor, your own) reads and writes the same memory; any LLM — Claude, GPT, Gemini, Llama, Mistral, DeepSeek — can drive that agent. Switch the model, keep the memory. Vector embeddings and the structured graph share one substrate.

Memory is the access layer on top of the Object Store, optimised for what AI agents need: a typed graph for structured recall, vector embeddings for semantic recall, and a stable MCP contract so the same memory survives a change of model. When the state-of-the-art LLM rotates next quarter, you change the runtime that calls the tools — the typed objects, the embeddings, and the integrations underneath stay exactly as they were.

Key features

• MCP-native — vendor-agnostic by construction
• Vector embeddings + structured graph in one place
• Sparse tensor row per node — same primitive as the model

How it works

Each node in the graph carries a sparse-tensor attribute row — the same primitive every AI model speaks natively — so the memory layer and the inference layer share one substrate instead of two adapters bolted together. Embeddings are computed in-process via the vector pipeline and stored alongside the structured attributes.

• Cross-model portability — why MCP makes the memory layer vendor-agnostic.
• Tensor substrate — the sparse-tensor row per node.
• Vectors — the embedding pipeline and similarity search.
• Senses — how upstream systems extend memory without copying data.

Try it

Point any MCP-aware agent at the IOBOXX MCP server and Memory starts working. See the MCP reference and the quickstart.