NNyquest docs

What Gets Remembered

The fact extractor decides what to remember from your conversations. This article explains what counts as a "fact" and what doesn't, with examples.

Quick mental model

The extractor looks for durable claims — things that stay true across conversations and could be useful later. It ignores transient, ephemeral, or one-off content.

Likely extractedLikely NOT extracted
"I'm a backend engineer working in Rust""What does this error mean?"
"My team uses GitHub Actions for CI""Can you summarize this article?"
"Patient is allergic to penicillin""Hello! How are you?"
"I prefer concise replies, no preamble""Thanks for the help!"
"We're targeting Q3 launch for v2""That's a good point"
"My partner's name is Alex""What's 2 + 2?"

The pattern: facts that describe you, your context, your preferences, your work, or named entities you care about stick. Generic Q&A doesn't.

Examples by category

Preferences and style

  • "User prefers Rust over Go for systems work"
  • "User wants 200-word maximum replies unless asked otherwise"
  • "User uses they/them pronouns"

Ongoing context

  • "User is studying for the bar exam, target date June 2026"
  • "User is a Series B startup founder, ~30 employees"
  • "User is renovating their kitchen, $40K budget"

Decisions made

  • "User decided to migrate from MongoDB to Postgres in Q1 2026"
  • "User chose Anthropic Sonnet over GPT-5 for the customer-facing assistant"

Technical setup

  • "Codebase uses Rust 2024 edition, axum 0.8, sqlx for DB"
  • "Frontend: React 19 + Vite + Tailwind v4"
  • "Deployed on Hetzner cloud, single VPS"

People and entities

  • "Sarah is the user's business partner at Acme Co"
  • "Patient JS is a 67yo male with T2DM and CKD stage 3"
  • "User's main client is named Acme Corp"

Pricing, dates, numbers worth remembering

  • "User's product currently priced at $29/mo"
  • "Project deadline is 2026-06-15"
  • "Annual revenue ~$2M ARR"

Things the extractor avoids

  • Transient pleasantries — "Thanks!" "Got it." "Good morning" — too small to remember
  • Single-turn questions — "What's the capital of France?" — not about you
  • Code snippets — useful in context but not as standalone facts (saved as artifacts instead)
  • Hypotheticals — "If I were to switch databases..." — not yet a decision
  • Quoted material — when you paste an article, the article's content isn't extracted as your facts

Fact lifecycle

A fact, once extracted:

  1. Active — used for recall, returned in queries
  2. Used — when recall surfaces it, last_used_at updates
  3. Superseded — if a newer fact contradicts it ("user moved to Berlin" supersedes "user lives in NYC")
  4. Retired — automatic housekeeping, no action from you. A daily consolidation pass merges duplicate paraphrases of the same fact into one canonical entry and retires transient ones; a separate pass retires facts that go long enough without being recalled.
  5. Forgotten — you remove a single fact yourself, in Settings → What Nyquest remembers. That fact is gone permanently.

Superseded and retired facts are excluded from recall but kept on record for audit trails.

Retirement is the usual explanation for memory that seems to vanish on its own: nothing was deleted, the fact either aged out or was folded into a duplicate. Pinned facts are exempt from both passes — pin anything you want kept indefinitely.

Fact confidence

Every fact carries a confidence score, shown next to it in Settings → What Nyquest remembers:

  • It rises when you re-affirm the fact — say the same thing again in a later conversation and the score goes up
  • It falls when you contradict it
  • It feeds recall ranking, so a higher-confidence fact wins against an equally-similar one

Contradicting a wrong fact is the gentle correction. Forget is the immediate one.

What if the extractor gets it wrong?

Two failure modes:

  • Extracted something private or wrong. This is rare but real. The extractor is conservative but not perfect. Open Settings → What Nyquest remembers, find the fact, and hit Forget — that single fact is removed permanently and nothing else is touched.
  • Missed something important. Just say it again in a future conversation. The extractor is more confident with repeated mentions, and re-affirming also raises the confidence of a fact that already exists.

Per-fact controls

Settings → What Nyquest remembers lists your facts with their confidence. For each one:

  • Forget — removes that one fact permanently
  • Pin — the fact never decays, is exempt from the daily consolidation pass, and is prioritised in recall

Recall itself has no on/off switch.

Coming:

  • Per-project memory scope (don't mix work + personal)
  • Tagging facts for explicit recall control

Where to next