n8n and Cognee explain how to give AI agent workflows memory that lasts across runs
In a guest post on the n8n blog, Veljko Kovac, head of FDE at the memory platform Cognee, explains that an LLM remembers nothing between calls, so an agent's memory has to live outside the model. The post separates short-term memory (conversation history and task state) from long-term memory (user preferences, outcomes of past tasks and company knowledge), and explains how agent memory differs from RAG: the agent also writes new facts back instead of only reading. It compares vector databases, plain files, knowledge graphs and hybrid systems, and stresses that memories must be consolidated, updated and deliberately forgotten, not just accumulated. For n8n it shows two patterns: a shared company brain that collector workflows feed and that acting workflows query before they act, and a memory layer attached to the AI Agent node's Tool port alongside the built-in Simple Memory for the chat.
Otomatzia editorial · Published
Source published: · n8n
