Long-Term Memory

Unlike short-term memory (in-memory, temporary), long-term memory uses external databases to store information indefinitely.

Agents can remember past conversations, user preferences, and learned skills across sessions.

Use SQL databases (PostgreSQL, MySQL) for transactional data, user profiles, and data requiring ACID guarantees.

Document stores (MongoDB) and key-value stores (Redis) offer schema-less storage, ideal for rapid prototyping and horizontal scaling.

Hybrid architectures combine SQL (structured data), NoSQL (flexible storage), vector DBs (semantic search), and graph DBs (relationships).

Text is converted to high-dimensional vectors (e.g., 1536 dimensions) that represent meaning, enabling similarity search.

Unlike keyword search, vector databases find semantically similar content—"forgot password" matches "reset credentials".

Retrieval-Augmented Generation: (1) Search vector DB for relevant docs, (2) Inject into prompt, (3) Generate grounded answer.

Choose based on scale, budget, and infrastructure—Pinecone (managed), pgvector (Postgres extension), Chroma (local dev).

Vector search with embeddings—high precision for meaning-based queries, but requires embedding models.

Fast, simple keyword search—excellent for exact phrases and proper nouns, but misses synonyms.

Merge vector and keyword results (e.g., 70% dense, 30% sparse) to capture both semantic and exact matches.

First: Fast broad search (100 candidates). Second: Expensive reranker model selects top 5. Production standard for quality.

Filter by user_id, date, tags before searching—narrows search space, prevents cross-user leakage.

Cache common queries (semantic caching), result sets (TTL-based), and hot paths (precomputed answers).