Back to full roadmap
topiccore
Vector Memory (Semantic Retrieval)
Embed past interactions, retrieve by semantic similarity to the query.
3 hours1 prereqs
Instead of stuffing the entire conversation into context: embed each message (or N-message chunk), store in a vector DB. For each new query, retrieve top-K similar past interactions, inject into prompt.
Pro: unlimited "memory" — only relevant pieces enter context.
Con: semantic similarity ≠ relevance. Sometimes the wrong chunk comes back. Hybrid retrieval (BM25 + dense) + reranking mandatory.
Stack: pgvector (Postgres), Pinecone, Qdrant, Weaviate, Chroma.