Infrastructure · Vector DB
Vector Databases for AI Memory
Vector databases store and search memory embeddings at scale — the most common long-term memory backend for AI agents using semantic retrieval.
Vector role
Role
What vector databases do for agent memory
Vector DBs persist embeddings + metadata, run approximate nearest-neighbor (ANN) search at retrieve time and filter by user/session — they are storage, not the full memory pipeline.
Extraction, consolidation and orchestration live in memory frameworks (Engram, Mem0, Zep) or custom code. The vector DB is the cold tier where semantic memories live. Engram adds memory semantics on top of Weaviate — not just raw vector storage.
Options
Popular vector databases for agent memory
Memory-framed — what each offers for agent backends, not generic DB marketing.
| Database | Agent memory use case |
|---|---|
| Weaviate (+ Engram) | Unified vector + memory layer; hybrid search; managed cloud |
| Pinecone | Managed ANN; pair with Mem0 or DIY extraction |
| pgvector | Postgres extension; DIY memory on existing DB |
| Qdrant | Self-hosted or cloud ANN; metadata filtering |
| Milvus | Large-scale ANN; high-volume agent fleets |
| Chroma | Lightweight local/dev prototyping |
Disambiguation
Vector DB vs memory framework
| Layer | Provides | Examples |
|---|---|---|
| Vector DB | Storage + ANN search | Weaviate, Pinecone, pgvector |
| Memory framework | Extract + write + retrieve + merge | Engram, Mem0, Zep, Letta |
| DIY | You build the pipeline on a DB | pgvector + custom code |
Schema
Schema design for agent memories
Standard metadata fields enable filtering, ranking and multi-tenant isolation.
{
"user_id": "usr_48291",
"agent_id": "support-v2",
"content": "User prefers email support",
"embedding": [0.12, -0.34, ...],
"timestamp": "2026-07-02T14:32:00Z",
"memory_type": "semantic",
"source": "session_a8f2",
"importance": 0.9
}FAQ
Frequently asked questions
What is the best vector database for AI agent memory?
Depends on stack: Weaviate + Engram for unified memory+RAG; Pinecone/pgvector for DIY with Mem0; Qdrant/Milvus for self-hosted scale. See comparison table above.
Weaviate vs Pinecone for agent memory?
Weaviate offers hybrid search and Engram memory layer natively. Pinecone is managed ANN — pair with Mem0 or custom extraction. Engram fits Weaviate-native stacks.
Is pgvector enough for agent memory?
Yes for DIY — you build extraction, merge and retrieval on Postgres. Frameworks (Engram, Mem0) add the pipeline. Good when you already run Postgres.
Engram vs raw Weaviate for memory?
Engram adds memory extraction, pipelines and semantics on Weaviate — not just vector storage. Raw Weaviate requires you to build the full memory loop. See Engram explained.
Which backend does Mem0 use?
Mem0 abstracts the backend — vector store configurable (often Qdrant, Pinecone or pgvector under the hood). See Mem0 alternatives.
Multi-tenant schema for agent memories?
Always include user_id (and optionally tenant_id) as indexed metadata. Filter every query by user_id — never rely on application-layer filtering alone.
Vector database vs knowledge graph for memory?
Vectors for semantic similarity; graphs for relationships and temporal validity. Zep uses graphs; Engram/Mem0 use vectors. See vector vs knowledge graph.
Vector DB cost at scale for agent memory?
Cost = storage (embeddings) + query volume + embedding API calls. Selective retrieval (~1,800 tokens/query) beats full-context (~26,000) on LOCOMO (Chhikara et al., 2025) — reduces inference cost even if storage grows.