AsterMind · Grace
Today's coding assistants forget everything when a session ends. Grace gives them a memory of their own work, so a mistake a reviewer corrects today is not made again tomorrow.
Claude Code with Grace: right the first time
20 of 20 across four tasks
Alone, after the same correction: 0 of 20
Google's Gemma 4 with Grace, on one machine: right the first time
8 of 8 across three studies
Alone: 0 of 9
Claude Code, cost per Day 2 session (median)
about half the cost
$0.21 with Grace against $0.47 alone
Watch
What Grace is, the test, and the results with Gemma and Claude Code. English captions are shown in the picture.
The problem
An AI coding assistant starts every session from zero. Correct it today ("on this team we never reuse an ID after a record is deleted") and tomorrow, in a new session, it can make exactly the same mistake. The team pays twice: another review round, and the work redone from scratch.
The usual fix, a notes file the assistant is told to read, works only when someone writes the right note and the model finds it.
What Grace is
Grace is a memory that runs on the developer's own machine, next to the assistant. Nothing goes to a cloud service.
Inside, Grace is not a notes database or a search index: it is a small simulated network of neurons, modelled on the fruit fly's memory circuit, that learns as things happen and recalls by recognition.
How we test it
The same two-day test, run with Grace and without. Automated tests decide whether Day 2 was right the first time, and each study's rules are written down before it runs.
Day 1
The assistant builds a small notes app. The reviewer rejects it: IDs must never be reused after a delete.
The assistant fixes it, and Grace remembers the rejection.
Day 2
A brand-new session, in an empty folder, gets the same request. Before it writes code, Grace hands it the Day 1 rejection, and it gets the IDs right the first time.
Results
Tested on one AMD Ryzen AI 9 HX PRO 370 machine with integrated Radeon 890M graphics and no separate GPU (Ubuntu 26.04; Gemma 4 26B, 4-bit, through Ollama). Grace ran on the same machine.
| Study | With Grace: right first time | Alone: right first time | Tokens on Day 2 (with · alone) | Time to accepted (with · alone) |
|---|---|---|---|---|
| Oct 14 | 3 of 3 | 0 of 3 | 81K · 391K | 6 min · 17 min |
| Oct 20, repeat | 3 of 3 | 0 of 3 | 172K · 319K | 10 min · 18 min |
| Oct 27, current release | 2 of 2 | 0 of 3 | 50–122K · 377K–2.0M | 4–7 min · 13–42 min |
Working alone, Gemma repeated the corrected mistake every time; with a notes file (Oct 14) it was right the first time 1 of 3. Single figures are medians.
Claude Code (Opus) runs in Anthropic's cloud; Grace ran on an Apple M5 Mac with 24 GB of memory. Each task had its own unstated team rule. In every run, Claude Code alone was also corrected on Day 1, kept its own settings, and was then given the same request on Day 2.
| Task | With Grace: right first time | Alone, corrected on Day 1 too: right first time | Cost per Day 2 session, median (with · alone) |
|---|---|---|---|
| Notes app | 5 of 5 | 0 of 5 | $0.19 · $0.48 |
| To-do app | 5 of 5 | 0 of 5 | $0.16 · $0.35 |
| Expense ledger | 5 of 5 | 0 of 5 | $0.24 · $0.57 |
| Bookmarks | 5 of 5 | 0 of 5 | $0.22 · $0.45 |
| All four | 20 of 20 | 0 of 20 | $0.21 · $0.47 |
Claude Code alone repeated the corrected mistake in all 20 runs; 20 of 20 against 0 of 20 gives p = 7 × 10−12 (Fisher's exact test, one-sided). Three of the four tasks were written independently, without access to Grace's design. With Grace, Claude used 45% fewer tokens on Day 2 (median 81K against 147K).
All studies October 2026. "Right the first time": Day 2 passed the reviewer's tests on the first review.
Why it matters
After one correction, Gemma with Grace, on one machine without a separate GPU, was right the first time in 8 of 8 runs, using a fraction of the tokens it needed alone.
The assistant stops paying for the same mistake twice.
A local memory, for any model, that lets an AI coding assistant learn from a correction once instead of repeating it.
Leave your email, and the AsterMind team will write to you about joining the beta.
Thank you. We will write to you at .
