Memora
An MCP memory server that gives AI agents persistent, collective memory, absorbing new facts with deduplication and update tracking and retrieving relevant memories, open tasks and sources through a graph-backed store.
Memora is an MCP memory layer for AI agents, providing structured storage, semantic retrieval, graph relations and source-backed context that persists across sessions. Feeding facts in through its absorb function has a language model classify each new fact against what is already stored, as a duplicate, an update, a contradiction, a related fact or something new, skipping duplicates and linking or consolidating related facts, with a dry-run preview available before committing.
Updates supersede older knowledge rather than deleting it, and retrieval can follow that lineage to the current version, the latest version, or the full history, depending on how it is queried. A single topic digest call bundles relevant memories, open to-dos and issues, related edges and source identifiers into one response. Search combines vector embeddings, from TF-IDF, sentence-transformers or OpenAI, with full-text, date-range and tag filters, and documents are stored as searchable fragment trees of claims, plan items, references and risks.
It is written in Python, stores data in SQLite with optional cloud sync to S3, R2 or D1, and supports serving multiple separate databases from one process.