Home

polymnemo

A shared long-term memory across any LLM, over MCP.

polymnemo is a provider-agnostic, cross-LLM long-term memory exposed as an MCP server. Whatever client connects — Claude Desktop, Cursor, or a custom GPT/Gemini agent — writes and reads the same memory, embedded by the same local model, so recall stays consistent across providers.

Principles

• Plain text is the single source of truth. Vectors are just an index; the original text is always stored, readable, and exportable.
• Zero external / generative-LLM calls. Embeddings run locally and deterministically (fastembed / ONNX, no API key). No generative "memory tax".
• HTTP-native. Streamable HTTP at /mcp; stdio only as optional local debug.
• Managed DB + semantic retrieval from day one. Neon Postgres + pgvector.
• Three pluggable layers: Auth / Store / Retriever.
• Bounded results, client decides. limit / cursor pagination; the model judges whether it has enough.

Differentiation (vs mem0)

No generative-LLM tax · original text stays human-readable · shared by design.

Navigation

• Implementation Plan — tech stack, data model, tool surface, pluggable layers, roadmap, risks.
• Architecture — diagram and data flow.
• Roadmap — phases mapped to milestones and issues.

Work is tracked as GitHub Issues grouped by Milestones (one per phase).