Nives no longer keeps notes about what it has forgotten.

Ask it to forget something and the memory goes — but the asking was itself memorable, so Nives could store a new memory about the forgetting: "you no longer want your canary word remembered", "you confirmed deletion of the bedroom cooling automation". Each one is harmless on its own. They quietly accumulate, and they come back when you ask Nives what it remembers about you — where a note saying that something isn't remembered is a peculiar thing to be told.

Nives now stores nothing at all about a forget. Notes of this kind already in your memory are cleared by the routine tidy-up that runs shortly after the add-on starts, so they should disappear on their own once you've updated. Nothing for you to do.

The same fix landed in the open-source sister project, home-mind 0.16.2, which has the technical detail if you want it.

Also in this release: the README now shows what Nives actually looks like in use — teaching it something, recalling it in a brand new conversation, changing your mind, and the automation it built appearing in Home Assistant's own automation list.

Upgrade: through the add-on store as usual. No configuration changes.

Nives is an AI assistant for Home Assistant that remembers you — tell it once that 100 ppm is normal for your NOx sensor, or that the WLED strip is "the main kitchen light", and it still knows next week. It installs as a single add-on: the conversation server and the memory engine live in one container on your machine, and the memory database never leaves it.

New here? Add https://github.com/hoornet/nives as a repository in Settings → Add-ons → Add-on Store → ⋮ → Repositories. The README has the full picture.

Existing users get updates through the add-on store as usual — this release note is a summary of where 2.4.x has arrived, not something you need to act on. Every version is written up in the changelog.

What's landed recently

You can ask it to forget (2.4.16–2.4.18). "Forget that my canary word is bumblebee — it's honeybee now." It quotes the exact memory back, waits for your yes, then forgets the old one and keeps the new one. It only ever removes the memory you named; if several could match it asks which; and "don't forget to water the plants" is still a reminder, not a deletion. A request that arrives outside a real conversation — an AI Task, say — is declined rather than deleting something nobody could confirm.

It knows its own name, so you can change it (2.4.17, 2.4.19). Its name was never actually in its instructions — it was inferring "Nives" from the label it puts on automations, which meant a Custom Prompt giving it a different personality was competing with something the model saw repeatedly, and smaller models drifted back. It's now told its name outright, so "You are HAL 9000, the calm and precise computer from 2001" replaces it cleanly and stays replaced. Ask it to rename itself in conversation and it now points you at the field that actually works instead of trying and failing silently. Thanks to @pitzoid, whose #54 is responsible for most of this list.

It won't learn its author's name as yours (2.4.15). The instructions Nives uses to decide what's worth remembering carried example facts that used a real name — the developer's. Most models read those as formatting samples; a small local one can copy an example straight into your memory. Neutral placeholders now.

Automations you describe in words (2.4.5–2.4.12). Ask for one and Nives shows you exactly what it will build — including what it won't do, like a rule with no conditions running regardless of who's home — and creates it only after you agree. It handles requests that need two automations at once, warns about the fixed-time-plus-threshold trap, and says which fields an edit leaves untouched.

It no longer restarts Home Assistant on its own (2.4.11) — it leaves you a notification and lets you pick the moment.

One Cloud plan, one model, kept current (2.4.10). Nives Cloud is optional and nothing is locked behind it; BYOK is free and never touches our servers.

Also in the box

An AI Task provider (ai_task.generate_data, text or structured output, usable inside your own automations), camera-snapshot vision for low-false-alarm alerts, Voice PE mic reopening, and arm64 images so it runs on a Raspberry Pi.

Sister project

home-mind is the same engine the DIY way — Docker Compose, every choice yours, AGPL, and it got conversational forgetting at the same time this did. Nives is that stack as one self-configuring add-on. Both are maintained; pick by temperament.