NNyquest docs

Recall and Surfacing

Nyquest selects memory separately for each request. The topic, conversation, project, saved facts, and available context budget all affect what the model receives.

What selection does

Before responding, Nyquest combines recent conversation history and an available summary with eligible saved facts. For a suitable message, semantic search also finds related historical excerpts. Selection removes duplicates, respects scope, and fits whole memory items into a limited budget.

The number of search candidates can be larger than the number supplied to the answer. A relevant item can be omitted because it is out of scope, duplicated, unavailable, or outside the context budget. Pinning increases priority and protects against housekeeping; it does not guarantee recall.

Ask clearly about history

To look for an earlier discussion, say so and name the subject: “What did we decide about pricing in our previous conversation?” Historical messages and summaries are treated as historical context, including when they contain a prior assistant answer.

Project-specific material stays within its project. A global preference can apply across conversations, while a local task instruction stays within its narrower scope. An explicit historical question does not make another project's material eligible.

Short messages and delayed updates

With the default gate, messages shorter than four words skip a fresh semantic lookup. That does not clear the recent conversation, its available summary, or eligible saved facts. A full, specific question gives retrieval more useful context.

New memory is processed in the background after a reply. It may not be searchable immediately. Completed streamed assistant answers can be indexed; interrupted or cancelled answers remain in history without semantic indexing as completed answers.

If semantic search is unavailable, Nyquest can continue using the available scoped context. A successful answer therefore does not prove that a new search ran.

Read the Memory Context panel

Open the answer's details or Inspect view and find Memory Context, when available. Current source details show:

  • The distinct memory items supplied to that answer
  • How many candidates were considered
  • Whether a conversation summary was also supplied
  • Each item's text, type, scope, date, author role, and source message identifier when available
  • Whether the item came from a user statement or other unverified historical context

The supplied-item count and candidate count describe different stages. Neither is a confidence score for the answer. A source message identifier identifies provenance; it is not an independent factual citation.

Detailed source snapshots are available for recent answers for up to seven days. Older responses may have only partial legacy excerpts. If details have expired or cannot load, do not infer missing sources from the count alone.

Memory and factual sources

Memory can report what you or an assistant said previously. It does not establish that an external claim is accurate today. Summaries can omit details, and older statements can be wrong or outdated.

For current factual questions, check the original evidence or use current sources. When web grounding is present, its linked sources appear separately under Web Grounding in Inspect. A remembered statement and a live web source serve different purposes; inspect the source that supports the claim you need to verify.

Improve the next reply

  • Name the project, decision, or topic in your question.
  • Reopen the original conversation when exact wording matters.
  • Include essential constraints directly in the current message.
  • Review Settings → What Nyquest remembers to forget stale facts or pin useful ones.

Browser chat has no global recall toggle. PAT chat-completion requests skip this memory layer. See Memory Privacy and Deletion for clearing stored memory.

Where to next