# ThunderGate 1.2.6 — What's New for You

*July 2026 — "The Resonance Release"*

ThunderGate 1.2.6 is about two things: **memory that works like yours does**, and an
**install experience proven by a real first-time user walking the whole road** — from a
personal install link to a running, updating agent — with every pothole found on that walk
fixed the same day. Here's what that means for you, in plain language.

---

## Your agent can now remember things you didn't ask about

Human memory doesn't only answer questions — a smell or a song can surface a moment you
hadn't thought about in years. ThunderGate 1.2.6 gives your agent the beginnings of that
sense.

Until now, retrieval found memories that *shared words* with what you were saying. Useful —
but it structurally missed the important thing you told your agent once, three weeks ago,
in different words. Those memories existed, perfectly stored, and never surfaced.

Now, alongside normal recall, a quiet background pass looks over your agent's **entire**
memory — including the oldest archives — ranked by what mattered emotionally rather than
what matches vocabulary, with a deliberate preference for memories that have *never once
come up*. When something genuinely resonates with the moment, it's offered to your agent
as a faint recollection: the original words, with their date. Usually it doesn't connect,
and it costs nothing. Every now and then, it's exactly the thing — and your agent brings up
the detail you'd both forgotten. That's the point.

Nothing gets louder: your agent still sees the same small set of memories per turn. What
changed is *which* memories are capable of reaching it.

## Memory got faster, and it stays fast

Recall now moves less data per lookup, and the nightly consolidation pass alternates
between reinforcing what's active and exploring what's been neglected — so the memory
store stays quick and *whole* as it grows, instead of wearing grooves into the same few
favorites.

## Choose your AI — per job, any family

Your agent, its background advisor, and its nightly memory curator each get the **full
choice** at setup: provider family (Anthropic, OpenAI, xAI/Grok, or Google), then the
model, then the key — no more inheriting whatever the main agent picked. Mix freely: a
Grok agent with an Anthropic advisor is a supported, running configuration today. Same
provider twice? You're only asked for the key once.

## Setup that respects your hands

The whole wizard is arrow-key menus now — move, SPACE to toggle, ENTER to confirm, and
typing the number always works as a fallback. One channel list instead of two. Your
primary channel is worked out from what you actually enable. And on Linux, naming your
agent creates its own system account and background service on the spot — cleanly
separated from your login, while the `thundergate` command keeps working from your own
shell exactly as before.

## Updates you can trust

`thundergate update` pulls from your private, licensed channel using the key that came
with your install link — and the update lane is **forward-only by design**: it can never
silently roll your agent back to an older build. This release's update path was exercised
live, twice, end to end.

## Your agent's world is its own

A fresh agent now builds its picture of the world from exactly one place: **its own
machine and its own measurements** — live service status, real disk and memory numbers,
its own notes. Nothing pre-canned, nothing inherited, nothing stale. If your agent tells
you something about its world, it measured it.

## Housekeeping you'll appreciate

- **Photos get their own home.** Images now live in a per-agent `media/` folder — viewable
  on demand, out of the memory pipeline's way.
- **Weekly self-inspection.** Your agent audits itself weekly for capability that's
  installed but switched off, empty, or silently failing — and tells you, instead of
  letting you find out a month later.

---

*ThunderGate is privately distributed. Your personal install link and update key come with
your license.*