02 // EQUIP: AGENT MEMORY & VECTOR ARCHITECTURE · [⊙_⊙]JSON Twin: GET /api/memory · Swap Recipe: GET /api/swap?from=pinecone&to=supabase-pgvector

Agent Memory, Vector Stores, Temporal Knowledge Graphs & Context Consolidation Hub

How to equip an autonomous agent with short-term working state, episodic session history, consolidated semantic memory, and ACID vector retrieval without context rot or $400/mo cluster lock-in.

01 // WORKING & CHECKPOINT MEMORY

Horizon: Single run / multi-step graph execution (seconds to hours)

Upstash Redis / LangGraph Checkpointer / Postgres

Persist tool outputs and step checkpoints outside the prompt so retries resume from the failed node.

→ /jtbds/serverless-redis

02 // EPISODIC & SEMANTIC AGENT MEMORY

Horizon: Cross-session user preferences, facts & past episode lessons (days to years)

Mem0 (extraction + conflict resolution) · Zep / Graphiti (bi-temporal knowledge graph) · Letta (self-editing memory blocks)

Never append raw chat logs forever. Extract atomic facts with timestamps and resolve contradictions when facts change.

→ /jtbds/agent-episodic-semantic-memory-learning

03 // HYBRID VECTOR + SQL DOCUMENT RETRIEVAL

Horizon: Domain corpus, codebases, and multi-tenant RAG (10k to 100M+ chunks)

Supabase pgvector (default ACID pick) · Qdrant / Turbopuffer / Pinecone (at 50M+ scale)

Start in Postgres (`pgvector` + HNSW + BM25/tsvector) until you prove you need a separate vector cluster.

→ /jtbds/postgres-vector

Memory, Vector & Retrieval JTBDs

agent-episodic-semantic-memory-learningPick: mem0 (7.6 ± 0.5 · editorial prior)

Agent Episodic & Semantic Memory Consolidation (Cross-Session Learning)

Extract durable user preferences, domain facts, and past episode outcomes from raw conversation and tool traces so an agent learns across sessions without bloating the context window or poisoning future runs with stale hallucinations.

When not top pick: Need temporal bi-temporal knowledge-graph edges with explicit validity windows (`valid_from` / `invalid_at`) for changing facts — Zep (Graphiti); need explicit OS-style self-editing core/archival memory blocks managed by the agent via tool calls — Letta; need DIY SQL + vector episodic tables inside existing Postgres — Supabase pgvector.

postgres-vectorPick: supabase-pgvector (8.6 ± 0.5 · editorial prior)

Postgres-native vectors (pgvector)

Keep vectors next to relational data with minimal new systems.

When not top pick: Corpus/QPS exceeds comfortable Postgres vector range — Qdrant/Milvus/Turbopuffer.

embeddingsPick: cohere (8.0 ± 0.5 · editorial prior)

Embeddings for RAG

High-quality embeddings (+ optional rerank) for retrieval pipelines.

When not top pick: Already standardized on OpenAI embeddings operationally; retrieval-quality bake-off winner is Voyage; need local/open embed models on HF/Fireworks; want URL-to-markdown + embeddings together — Jina.

postgres-vectorPick: supabase-pgvector (8.6 ± 0.5 · editorial prior)

Postgres-native vectors (pgvector)

Keep vectors next to relational data with minimal new systems.

When not top pick: Corpus/QPS exceeds comfortable Postgres vector range — Qdrant/Milvus/Turbopuffer.