How Memory Works

Nyquest can bring relevant context from your conversations into signed-in browser chat. Memory helps with continuity, but it does not preserve every detail or guarantee that a past statement will be recalled. PAT chat-completion requests are stateless and skip this memory layer.

What contributes to a reply

Long conversations are shortened to fit the context budget. A saved transcript can contain more than the model receives for any one reply.

Memory updates after a reply

Fact extraction, summary updates, and search indexing run in the background after the response. A useful detail may take time to become available, and an exchange may produce no new facts.

The extractor requires supporting text from your message. It checks the result before saving it and rejects malformed, unsupported, or incomplete results. A failed extraction can leave a gap in memory; it does not mean your chat transcript was deleted. Historical failed extractions are not automatically replayed.

Completed streamed assistant answers can become searchable historical context. Interrupted or cancelled answers stay in history without being added to semantic recall as completed answers. An assistant's earlier conclusion remains an unverified historical claim.

Scope matters

Memories can apply to the original conversation, its project, or across conversations. An explicit lasting preference such as “I prefer concise replies across conversations” can be treated differently from “keep this answer short.” Temporary task instructions and quoted instructions should not become global preferences.

Project-specific context is kept within its project. Global preferences may still apply there. Asking about a previous conversation can make historical excerpts eligible, but does not cross another project's boundary or guarantee a match.

When retrieval runs

Nyquest assembles available conversation context before responding. With the default short-message gate, messages under four words skip a new semantic search. Recent history, an available summary, and eligible saved facts can still contribute.

For a historical question, give the topic and make your intent clear: “What did we decide about the database in our previous conversation?” A specific question works better than “remember that?”

See what was supplied

Open the answer's details or Inspect view, then Memory Context, when available. It shows the distinct memory items supplied to that answer, their source information, and whether a conversation summary was also supplied. Details for older answers may be incomplete or unavailable.

A memory count describes context supplied to the model. It is not a measure of truth or proof that the answer used every item. See Recall and Surfacing.

Manage saved facts

In Settings → What Nyquest remembers:

• Pin protects a fact from automatic decay and consolidation and gives it priority in selection. It does not guarantee inclusion in every answer or change its scope.
• Forget removes that stored fact from future recall. It does not erase the original conversation.

Automatic housekeeping can merge duplicate facts and retire unpinned ones. Repeating a detail does not guarantee that it will be saved or selected. For critical work, include the exact context in your current message or reopen the original conversation.

Signed-in browser chat has no global recall toggle. To clear stored memory, use Settings → Danger Zone → Delete Memory & Knowledge. See Memory Privacy and Deletion for the difference between clearing memory and deleting chat history.

Where to next

• What Gets Remembered — examples and corrections
• Recall and Surfacing — selection and source details
• Memory Privacy and Deletion — storage and controls
• Conversation and Memory Issues — troubleshooting