Memory
By the end of this page you will know how to enable each memory tier so your agent can remember preferences, recall past sessions, and draw on a knowledge base.
Prerequisites
- An AI Agent open in the editor (Create an AI Agent).
- Admin or Developer role.
- For Episodic and Semantic memory: a Content Store (project default or a specific store). See RAG & knowledge.
How memory works
The Memory tab adds a memory layer that improves context awareness and decision-making from past experience. It has three tiers, each opt-in — every toggle defaults to off. Each tier has separate read and write controls, so you can let the agent recall without writing, or vice versa.
Profile Memory
Cross-agent user preferences — locale, timezone, communication style. Stored in the project database, so no Content Store is needed.
Enable profile access
Turn on Read user profile to let the agent use a user’s preferences, and Write / learn preferences to let it learn and store them.
Episodic Memory
Session summaries stored as RAG chunks, so the agent can recall context across multiple sessions.
Enable episodic memory
Turn on Recall past conversations and/or Save session summaries.
Choose a Content Store
When either toggle is on, pick the Content Store used for episodic summaries — a specific store, or the project-level default.
If no Content Store is selected, episodic memory won’t function. Select a store above (create one first under Content Store if you don’t have one).
Semantic Memory (Knowledge)
Self-learned knowledge plus linked Knowledge Bases for org documents and policies.
Enable semantic memory
Turn on Search knowledge base to let the agent retrieve knowledge, and Enable self-learning to let it extract and store new knowledge from conversations.
Choose a Semantic Content Store
When either toggle is on, select the Semantic Content Store — used both for searching and for storing self-learned knowledge (the same project-default rules and warning apply as for episodic memory).
(Optional) Override the extraction prompt
When self-learning is on, a Semantic Extraction Prompt Override field appears. Use it to customize how the agent extracts knowledge (up to 12,000 characters). You can view the built-in default and reset to it at any time.
Memory read/write settings become tools the agent can call at run time (for example, profile read/write and knowledge search). Enabling a tier is what makes those capabilities available.
Save your changes
Click Save in the header to persist your memory configuration into a draft version.
Common issues
- “No Content Store configured” warning — Episodic or Semantic memory is enabled but has no store. Pick one, or create a Content Store first.
- The agent doesn’t remember across sessions — Episodic Save session summaries must be on (summaries are written when a session closes), and Recall past conversations must be on to use them.
- Self-learning isn’t storing anything — Ensure Enable self-learning (write) is on and a Semantic Content Store is selected.