Agent Memory lets an agent learn from past conversations, carrying facts, events, and learned procedures forward instead of starting fresh every time.
Overview
Memory is described in-product as: "Configure how your agent learns from past conversations." You configure memory with memory profiles (each defines a kind of memory) and memory sets (bundles of profiles you can reuse across agents). Memory is off by default, so an agent only remembers once you turn it on.
Turning On Memory
Use the Enable Memory toggle — "Activate memory storing capabilities for your AI agent." It's off by default; the memory profiles and memory sets sections only appear once it's on.
Note: Enable memory for agents that benefit from continuity — a personal assistant, a support agent that should recall a customer's history — and leave it off for stateless, transactional agents.
Adding a Memory Profile
Use Add Memory Profile to define one kind of memory. A profile has:
Field | Description |
|---|---|
Memory Type | Semantic ("Factual Knowledge"), Episodic ("Event History"), and Procedural ("Skills & Rules"); pick the types to retain. |
Scope | User Specific (private per user), Team Specific (shared within a team), or Agent Specific (shared across all users). A platform-provided Project-scoped profile can also appear, read-only — it isn't a scope you choose yourself. |
Generation Settings | The model and instructions that decide what's worth remembering, plus the embedding model and refresh frequency. |
Storage Settings | An optional expiration period so memories age out. |
Clear instructions on the profile matter — they tell the agent what to store and what to ignore.
Each profile you've added appears as a card, and its ⋮ menu offers View Memories (open what the agent has actually remembered under this profile), Edit, and Delete. Deleting asks you to confirm, because it can't be undone. The menu is hidden entirely when the section is read-only — for example while you're viewing a published version.
Reusing Memory Across Agents
Use the Memory Sets section to attach a memory set — a reusable bundle of memory profiles — to the agent with Add Memory Set. Build a memory set once and link it to several agents so they share the same memory configuration. If none are attached, the section shows an Add Memory Set row. If the picker reads No Memory Sets Found, create one first under Memory Sets.
Note: Use profiles for memory specific to this agent, and memory sets when several agents should remember the same way.
How Memories Reach the Agent's Replies
When memory is enabled, retrieval happens at the start of each request, before the model is called:
The platform resolves which memory profiles apply — the agent's own profiles and attached sets, sorted by scope (user / team / agent) — skipping anything the user has opted out of and any team profile the user isn't a member of.
The user's message is embedded and matched against stored memories by meaning (vector search), pulling a shortlist of candidates (up to about 20).
A strict relevance filter — a small model call instructed to keep only memories that genuinely help answer the query — trims the shortlist, and only the survivors are injected into the agent's system prompt as prior context. Language-preference memories ("always answer in French") are handled specially so they reliably reach the agent.
The agent never sees the whole memory store — it sees a few memories judged relevant to the current message. That keeps prompts small and answers on-topic, and it means an irrelevant stored memory quietly stays out of the way.
When the Agent Actually Remembers Something New
Writing works on a schedule, not per message: each memory profile syncs at its Refresh Frequency, sweeping the conversations that changed since its last run, extracting facts from what users said, and de-duplicating them against existing memories before storing.
Note: Expect a short lag between a user stating a fact and the agent recalling it in a later session — up to one refresh interval. Individual users can be opted out of memory entirely, in which case nothing is written from (or read into) their chats.