Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.corti.ai/llms.txt

Use this file to discover all available pages before exploring further.

[
  {
    name: "memory-expert",
    type: "reference",
  }
]
Memory is a core toolbox expert — it is foundational to most agent workflows where the orchestrator needs to reason over large inputs without bloating the prompt.
The Memory expert lets agents recall and analyze content from large contexts — long documents, attached files, transcripts, or other bulky inputs — that would not fit cleanly inside the orchestrator’s working prompt. The orchestrator hands the bulky context to the Memory expert and asks targeted questions; the Memory expert searches and analyzes that content and returns the relevant facts.
The Memory expert does not persist information across sessions, conversations, or context boundaries. It only operates over the context that has already been provided to the agent in the current request. If you need durable, cross-session memory, that has to be implemented at the application layer.
Use the Memory expert when a request involves a large body of text or files that the orchestrator should not load into its own prompt — for example a long discharge summary, a multi-page guideline, or an attached transcript — and you only need a few targeted facts back.

​
Capabilities

The Memory expert can:
  • Search large bodies of context for information matching a query
  • Analyze long documents or files and return targeted facts
  • Surface relevant excerpts so the orchestrator can ground its synthesis
  • Reduce the size of the orchestrator’s working prompt by offloading bulky lookups

​
Use cases

  • Querying a long clinical document or attached file for specific facts
  • Pulling targeted passages out of a large transcript or note
  • Summarizing or extracting from inputs that are too large to keep in the orchestrator prompt
  • Providing analyzed excerpts as evidence for a downstream expert

​
How to invoke

Send the relevant query or high-level objective as the text parameter. Do not pass new raw facts to the Memory expert — provide only the question or objective so it can search and analyze the existing in-request context itself. For questions that can be answered directly from the orchestrator’s own prompt without consulting bulky context, you can answer directly without invoking the Memory expert.