Summary
Memory with entity resolution, so a vague reference two weeks later still lands.
Most memory servers store text and search it. The part that matters here is entity_resolve and context_inject working together: you say 'that project', and the pronoun is resolved to a named thing before the search runs.
What it is
A memory server backed by Azure Cosmos DB and OpenAI embeddings. It stores what you discuss, extracts the people, companies and projects mentioned, and pulls the relevant pieces back when a later question refers to them.
What you get
- memory_store — save information with the entities pulled out of it automatically
- memory_search — semantic search across everything stored
- context_inject — fetch the context relevant to the question being asked right now
- entity_resolve — turn 'he' or 'that company' into the actual person or company
- conversation_analyze — patterns across your discussions
- memory_timeline — when a given entity came up, in order
Requirements
Node.js 18+, an Azure Cosmos DB account and an OpenAI API key. Cosmos endpoint, key, database and container names go in the environment, along with OPENAI_API_KEY. Built from source and run as a node process.