NNyquest docs

Conversation and Memory Issues

When conversations behave unexpectedly, history goes missing, or memory recall feels off.

A conversation disappeared

Conversations don't disappear on their own. If one is missing:

  • Wrong project filter. If you're inside a project, the sidebar shows ONLY that project's conversations. Click the × next to the project name in the chat header to return to top-level and see all conversations.
  • Archived. Right-click the conversation list → Archived view (or Settings → archive) to restore.
  • Deleted. Permanent deletion is exactly that — the conversation is removed from the live database. There's no recovery.
  • On a different account. Sign in to the correct account. Conversations don't sync across accounts.

If you're sure a conversation existed and it's gone (and you didn't delete it), talk to a real human — we can check our logs for unusual activity.

Search isn't finding a conversation

Search is currently fuzzy substring match (semantic search is Phase 6). It finds conversations where the query appears verbatim in your messages or the model's replies.

If a search misses:

  • Try shorter, more distinctive keywords — multi-word queries fragment
  • Search is case-insensitive but exact-ish on character match — "TypeScript" finds it but "type script" might not
  • Search ignores stop words ("the", "and", etc.) — your distinctive nouns matter most
  • Conversations from over 90 days ago might not appear in the live index — older ones are still searchable but may take longer

"I asked the model X yesterday and it doesn't remember"

By default, the model only sees the current conversation's history. Across separate conversations, the model has no memory.

Memory recall (the platform's auto-recall feature) does pull relevant past chunks, but it's biased toward recent and topically-similar conversations. If the model isn't recalling something:

  • Your message was too short. Messages under 4 words skip semantic recall — "and yesterday?", "my stack?", "do it" pull nothing new in. The recent messages, the conversation summary and your durable facts are still provided, but no search runs. Ask in a full sentence. This is the most common cause.
  • The previous conversation might have been deleted (conversation recall is removed; durable facts may remain until forgotten or deleted separately)
  • The previous conversation might not be similar enough — recall is semantic, not keyword
  • The fact was retired automatically. A daily consolidation pass merges duplicate facts and retires transient ones, and facts that go long enough without being recalled are retired too. Nobody deleted it — it aged out. Check Settings → What Nyquest remembers, and Pin anything you want kept: pinned facts never decay and are exempt from consolidation.

Workaround: start a new conversation by quickly summarizing the relevant prior context: "Earlier we were working on X. The constraints were Y and Z. Now I want to..."

This costs a small number of input tokens but works reliably.

Memory recall is surfacing irrelevant chunks

The memory layer recalls chunks based on semantic similarity to your current message. If irrelevant chunks appear:

  • The current message's keywords are accidentally matching an old unrelated conversation
  • Try a more specific phrasing of the current message
  • If one stale fact keeps surfacing, open Settings → What Nyquest remembers and hit Forget on it — that removes that single fact permanently
  • Contradicting the fact in conversation is the softer fix: its confidence score drops and recall ranks it lower

If the recalled chunks are wildly off-topic (suggesting a model confusion), talk to a real human — we can investigate the recall pipeline for your account.

Branching a conversation produces unexpected results

Branching forks the conversation at a specific message. The new branch:

  • Inherits the conversation's full history up to and including the branch point
  • Doesn't inherit later messages in the original conversation
  • Doesn't auto-rename — the new branch shows "(branched)" in the title; rename to clarify
  • Doesn't fork artifacts — artifacts are shared between original and branch

If a branch behaves identically to the original, you may have branched at the wrong message. The branch point matters; double-check which message has the edit pencil.

Project's system prompt isn't being used

If you're inside a project but messages feel un-influenced by the system prompt:

  1. Verify the project is actually active. The chat header should show the project name.
  2. Verify the system prompt is set. Project header → gear icon → check the System Prompt field.
  3. The system prompt is prepended to future messages, not retroactively. Existing messages were sent without it; new ones include it.
  4. If preferred model is set, verify it matches an available model. If preferred model is unavailable (BYOK key removed, etc.), the request falls back to your global model and the prompt still applies.

Conversation title is wrong

Titles auto-generate from your first message after the first reply completes. If the title feels off:

  • Right-click the conversation → Rename — set whatever title you want
  • The title is just a label; it doesn't affect the model's context

Where to next