Conversation memory

Within one conversation, OpenFeather keeps the recent turns, the resolved subject of the question, and any structured result already established — so "only Q3" is a complete instruction.

• The topic and entities of the current thread.
• Results already retrieved, so the provider is not called twice for the same thing.
• Constraints you stated earlier — date range, workspace, preferred format.

• Anything from a different conversation.
• Anything from a workspace you are not currently in.
• Results whose underlying data has since changed; those are re-fetched when a new run requires them.

• The subject changed and you want an independent answer.
• A previous run was cancelled or errored and you want a clean retry.
• You are comparing two scenarios and do not want them blended.
• Context has grown long enough that the topic is no longer what you care about.

Short follow-ups work best when the earlier turn was itself specific. If a follow-up produces a vague answer, restate the constraint instead of repeating the whole question: "same as above, but only spreadsheets" beats a full re-ask that changes one word.

Conversation context is bounded. As a thread grows, the earliest turns may fall outside the retained window while the current topic and recent results stay. A very long thread therefore gets slower and more expensive before it gets worse.

1. 1Notice when follow-ups start ignoring earlier detail.
2. 2Summarise the agreed conclusion into a durable memory entry.
3. 3Start a new conversation for the next topic.