Blog · July 21, 2026
Most AI "memory" is a diary: everything you ever said, appended forever. That sounds like a feature until you change your mind. Tell an assistant in March that you're deploying on one cloud provider, correct yourself in June, and a diary-style memory happily hands the model both facts — the truth and the thing you explicitly un-decided — and lets it guess. The more you use it, the noisier it gets.
Nyquest has had memory for a while: durable facts extracted from your conversations, semantic recall over everything you've discussed, and a token budget that keeps injected context lean. This week we rebuilt the layer underneath it so memory behaves like a belief state — a small set of things currently held true — instead of an append-only log. Here's what changed.
When Nyquest extracts a new fact from a conversation, the extractor now sees what it already knows about you. If the new fact updates or contradicts an old one — you changed a decision, finished a task, moved a deadline — the old fact is superseded: retired from recall, with an audit link recording what replaced it and why. The model gets the current truth, not a debate between versions of you.
Some facts get an even stronger guarantee. When a fact is a single-valued property of something — the database your project uses, the timezone you work in — Nyquest keys it as an entity property, and replacement becomes mechanical: a new value for the same key automatically retires the old one, no language-model judgment required. In our tests, "our primary database will be MySQL" followed later by "we've switched to PostgreSQL" leaves exactly one active fact standing, every time.
Even with good hygiene at write time, memory accumulates clutter: the same preference phrased three ways, tasks that were finished weeks ago, notes that mattered for exactly one afternoon. Once a day, a consolidation pass reviews each user's active facts and proposes conservative cleanups — merge paraphrases into one canonical fact, supersede what's been outdated, retire what's transient. It runs under strict guardrails: anything you've pinned is untouchable, every operation is soft (nothing is ever hard-deleted by the system), and every run is logged and reviewable.
On our own heavily-used development account, two passes compressed 138 accumulated facts down to 58 clean, canonical ones — and reading the merged results is the satisfying part: each one preserves the specific details from its sources without inventing anything new.
Every fact now carries a confidence score that learns from how your conversations treat it. Re-affirm a fact — build on it, rely on it — and its confidence rises. Cast doubt on it without replacing it ("actually, I'm not so sure about that anymore") and its confidence drops, without the fact being deleted: doubt is not a correction, and the system knows the difference. Confidence feeds directly into recall ranking, so the memories the model leans on hardest are the ones you've most consistently stood behind. Every adjustment is recorded, with before-and-after values.
Not every message needs the memory system. "Continue." "Ok, thanks." "Make it shorter." These lean entirely on the conversation you're already in — searching long-term memory for them adds latency and prompt noise for zero benefit. Nyquest now skips memory recall on trivial messages entirely. It's a small change with a pleasant effect: short follow-ups get snappier, and your context budget is spent where it matters.
In the spirit of honest numbers: while building this, we found that extracted facts had never been entering the semantic search index at all — the indexing function existed, but nothing called it. Facts that were never recalled looked permanently "cold" to our decay system, which dutifully archived nearly all of them. A memory system quietly forgetting because of a wiring bug is exactly the kind of failure you only catch by instrumenting everything. It's fixed, the stranded facts were recovered and indexed, and a self-healing check now runs at every boot so the index can never silently drift out of sync again.
Memory you can't inspect is memory you can't trust. Settings now has a "What Nyquest remembers" panel: every active fact, its confidence, and where it's keyed — with three controls that do exactly what they say. Pin makes a fact permanent and immune to every automatic process above. Forget removes a fact from recall permanently, effective immediately. And the existing danger-zone option to erase your memory entirely is still there. Your memory, your call.
The numbers behind this post
No account needed to start chatting — or bring your own API key.