Memory lets an agent remember durable facts, preferences and project context across messages and conversations. In AgentRow it is explicit: a read node before the model answers, a write node after, scopes that say who the memory belongs to, and an approval step before anything is saved.
Conversation history is what the model sees within one chat. Memory is what survives beyond it: a user's preferences, a fact about their account, the context of a project, an instruction they gave once. Without memory an agent starts every conversation from zero; with unscoped memory it leaks one user's context into another's.
From the agent's Memory tab: off, remember in this chat (scoped to the conversation), remember for this agent (shared by everyone talking to it), or remember this user (scoped to the user or end user across their conversations). Workflow nodes can address the user, conversation, agent, app and workspace scopes, and a namespace separates memory sets within a scope.
Signed-in end users get per-user memory through the app's identity token; anonymous visitors keep their memory with their browser session until they sign in and their history is merged.
Before the model answers, the memory read node searches for relevant items (recent, keyword, hybrid, importance, vector or semantic hybrid strategies, with top k and a minimum score) and injects them into the prompt under a relevant-memory heading with their source and score.
After the answer, the memory write node extracts candidate items (preferences, facts, project context, instructions) from the exchange. The default write policy is that candidates require approval: they appear in the memory items list where a team member approves or rejects them. Automatic saving above a confidence threshold can be enabled per agent. Sensitive content is excluded by default, duplicates are merged, and items can update existing ones or expire after a retention period.
Not by default. Candidates wait for approval in the memory items list. You can enable automatic saving above a confidence threshold per agent when you trust the extraction.
Set the agent to remember this user and let the app identify users with a signed identity token. Memory is then scoped to that user across all of their conversations and devices.
Yes. Each run records the memory read with the items considered and selected, and the generation shows the memory context exactly as it appeared in the prompt.
Ground AI agents in your documents: knowledge bases with chunking, embeddings, hybrid keyword and vector retrieval, reranking, citations and full logs.
Orchestrate several specialised AI agents behind one app: a router picks the agent, nested runs stay fully traced, and one answer streams to your users.
See how this feature shows up in a real agent app.
Build a Shopify AI agent: track orders and check inventory through the Admin API and answer product questions from a knowledge base, in one chat widget.
Join the waitlist for early access, or read how the rest of the platform fits together.
