> ## Documentation Index
> Fetch the complete documentation index at: https://docs.autosage.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Memory
> How AutoSage remembers context — within a conversation and durably across conversations.
AutoSage remembers context at two levels. They work together, but they're easy to conflate, so it helps to keep them straight from the start.
Recent turns remain verbatim, while older turns are summarized on a best-effort basis so long conversations stay useful within the model's context window.
Facts, preferences, and decisions the assistant recalls in future, separate conversations. Optional, and configured per tenant and per knowledge base.
## Conversation memory
Within a single chat, the assistant uses recent messages and, for longer conversations, a summary of older turns. You don't configure this; it's how every conversation works.
As a conversation grows, AutoSage automatically summarizes earlier turns to preserve useful continuity without slowing down. Summarization is best-effort and may omit details, so start a new chat for unrelated work and restate any critical constraints when needed.
Conversation memory lives inside one chat. When that chat ends, its context ends with it — unless durable memory is enabled to carry specific facts forward. Learn more in [Long conversations](/chat/long-conversations).
## Durable memory
Durable memory is what most people mean when they say "memory": the assistant remembering something from one conversation and recalling it in a completely different one, days or weeks later. A stated preference, a decision, a fact about the user or their work — captured once and available in future chats.
Durable memory is **optional and off by default**. You turn it on and tune how it behaves per tenant, and you can further shape it per knowledge base. See [Durable memory](/memory/durable-memory) for setup.
### Who a memory is scoped to
When durable memory is enabled, its scope depends on how the assistant is being used:
| Usage | Memory scope |
| -------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
| Dashboard users | Shared across the tenant, so the team builds a common memory |
| API-served end users | Isolated per end user, keyed to an external subject id, so one customer never sees another's memories |
| A knowledge base set to keep memory isolated | Kept fully separate to that knowledge base |
The per-knowledge-base scope can only make memory **more** isolated than the tenant setting, never less. This keeps a knowledge base's memory contained when you need it to be. See [Memory modes](/knowledge-bases/memory-modes).
## Which one do I need?
That's conversation memory, and it's always on. Nothing to configure.
That's durable memory. Enable it on the tenant and choose how facts are captured. See [Durable memory](/memory/durable-memory).
Durable memory is automatically isolated per end user via their external subject id, so each customer's memories stay private to them.
## Next steps
Enable it on a tenant and choose how facts are captured.
View, search, and delete what the assistant remembers.
See how long conversations keep their context.
Further isolate memory for a specific knowledge base.