Glossary
The Memorg vocabulary.
A short, plain-language reference for the concepts behind the library — the hierarchy, the vector index, and how retrieval is scored.
- Session
- The top level of the Memorg hierarchy. A session belongs to a user and holds one or more conversations. The token budget for retrieval is set when the session is created.
- Conversation
- A grouping inside a session. A conversation holds topics and, through them, the exchanges that make up a dialogue.
- Topic
- A subdivision of a conversation that groups related exchanges, giving retrieval a natural scope narrower than a whole conversation.
- Exchange
- The smallest unit of conversational memory — a message pair between the user and the system. Each exchange is stored and embedded for semantic search.
- Memory item
- A typed, tagged unit of memory that is not tied to the chat hierarchy — a document, a note, or an arbitrary blob you want to store and recall.
- Embedding
- A vector representation of text produced by an OpenAI embedding model. Memorg embeds exchanges and memory items so they can be found by semantic similarity.
- Vector index
- The USearch nearest-neighbour index that stores embeddings and answers similarity queries. It lives in the same file as the SQLite store.
- USearch
- The vector search engine Memorg uses for fast approximate nearest-neighbour lookup. It is embedded in-process alongside SQLite — no separate vector database.
- Blended retrieval
- Memorg’s deterministic ranking that combines semantic similarity, recency, and importance into a single score, rather than relying on similarity alone or an LLM reranker.
- Importance
- A per-item signal in the retrieval blend that lets durable or high-value memories outrank passing details, independent of how similar or recent they are.
- Recency
- A signal in the retrieval blend that favours newer memory, so recent context is weighted alongside similarity and importance.
- Token budget
- A per-session limit on how much retrieved context is returned. Memorg trims results to fit the budget so the payload always fits the model window.
- MCP server
- A FastMCP server shipped with Memorg that exposes memory read/write to MCP-aware clients like Claude Desktop and Cursor, without bespoke integration code.
- Local-first
- Memorg runs in your process against a single SQLite file you own. No managed service is required to develop or ship.
See these concepts in context on how it works, the features, or in real use cases.