Skip to main content
Durable memory lets an assistant carry facts, preferences, and decisions from one conversation into future, separate ones. It’s optional and off by default. You enable and tune it with two tenant settings, then optionally scope it further per knowledge base.

The two tenant settings

1

Choose a mode

The mode decides whether durable memory is active at all.
ModeBehavior
DisabledNo durable memory. The assistant only remembers within each conversation.
DurableDurable memory is on. The assistant can recall facts across conversations.
2

Choose a retention level

When the mode is Durable, retention decides how new facts get into memory.
RetentionBehavior
NoneRecall only. The assistant uses what’s already stored but adds nothing new.
Explicit onlyStores a fact only when a user explicitly asks it to remember something. See Memory commands.
AutomaticAlso captures relevant facts from conversations on its own, in addition to explicit requests.

Mode × retention at a glance

The two settings combine to give you precise control over reading and writing memory:
ModeRetentionRecalls existing memories?Stores new memories?
DisabledNoNo
DurableNoneYesNo
DurableExplicit onlyYesOnly when explicitly asked
DurableAutomaticYesYes, automatically and on request

Scoping further with a knowledge base

Each knowledge base has its own memory modeDisabled, Subject, or KB-restricted. A knowledge base can only further restrict the tenant setting, never widen it.
If the tenant has durable memory turned off, a knowledge base cannot turn it on. The tenant setting is always the ceiling, and the knowledge base setting narrows from there.

What happens when memory is enabled

Once durable memory is active, recall and storage happen quietly in the background:
  • On each new message, relevant memories are automatically retrieved and added to the assistant’s context.
  • The assistant indicates when a memory was used, so it’s clear where recalled context came from.
  • New facts are stored according to your retention level.
Recall and storage are best-effort. A memory-service failure does not block a normal answer; if nothing relevant is found, the assistant simply answers without recalled context.

Choosing your settings

Start with Explicit only if you want users to stay in control of exactly what gets remembered. Choose Automatic for assistants that should learn preferences and context over time with no extra effort from the user.

Next steps

Memory commands

Store facts conversationally with natural phrasing.

Manage memories

View, search, and delete what’s been remembered.

Memory modes

Restrict memory for a specific knowledge base.

Memory overview

Revisit how the memory levels fit together.