As AI workflows become more agentic, what patterns are people using for shared memory, context management, and state persistence across multiple agents in n8n? Are you relying on vector databases, MCP servers, custom APIs, or something else?
1 Like
Welcome @Pawan_Jain! Here’s what I’m using across production setups:
Short-term / session memory: n8n’s built-in Window Buffer Memory node (Postgres-backed) is solid for per-conversation context. Key is partitioning by sessionId - I set that to {{$execution.id}} for one-shot runs or a user ID for persistent chat sessions.
Long-term / cross-session memory: Postgres table with a user_id, key, value, updated_at schema. Agent reads relevant rows at the start of each run via a Postgres node, then writes updates at the end. Simple and easy to inspect.
Semantic retrieval: For agents that need to recall from a large knowledge base, I add a Supabase or Qdrant vector store node as a tool. The agent calls it only when the query seems to need historical context - keeps token usage low.
MCP servers are still overkill for most n8n setups unless you’re sharing memory across many different systems outside n8n. For most multi-agent workflows inside n8n, shared Postgres is simpler and reliable enough.