semantic-memory

Semantic search over project memory uses vector embeddings and cosine similarity to match natural language queries with stored architectural decisions and notes, retrieving relevant context based on meaning rather than exact keywords.

To retrieve context from architectural decisions and postmortems, query the project memory using natural language. The system applies vector embeddings to find and return related concepts, lessons learned, and past observations.

Yes, you need a local or API-based embedding model to generate vector embeddings. The system calculates cosine similarity against these embeddings, storing the resulting index data in a lightweight JSON format for retrieval.

The best way to search lessons learned without exact keywords is using semantic search. It encodes your natural language query into a vector embedding and retrieves matching project notes based on conceptual similarity.

Semantic memory stores vector index data in a lightweight JSON format. This approach allows the information retrieval process to access agent notes and project history without requiring a heavy external database.