← All rituals

Ritual iii · Reflect Grove

A journal that prompts,
never preaches.

Reflect Grove is a 5–15 minute reflective practice with a private, generative AI companion. The companion offers prompts, asks gentle follow-up questions, and never gives advice. It runs entirely on your device. We cannot read what you write. We have not built the infrastructure to.

Three prompts the companion might offer

Generated locally, never the same twice, calibrated to the texture of your last few entries. You can refuse any prompt — refusal is itself a signal the model learns from.

"What part of today, no matter how small, felt like it belonged to you?"
"There is a thing you keep nearly writing about. Would now be a moment to write a single sentence toward it?"
"Describe the room you are in, without using any adjectives. Notice what changes."
"What did you postpone today that, on reflection, you are quietly glad you postponed?"

Model card

Reflect Grove ships a distilled small-language-model finetuned on a curated reflective corpus. Below is the model card as of the most recent release. The model is shipped inside the application bundle; nothing is downloaded at runtime.

RG-Reflect-4B

v0.7 · released Apr 2026
Architecture
Decoder-only transformer, 4.0B parameters, 4-bit weight-only quantised on device.
Base
Open-weights base · custom continued pre-training on UK English contemplative corpora · supervised finetuning on a hand-curated 11,400-prompt reflective dataset.
Runtime
Apple Neural Engine · Qualcomm Hexagon · CPU fallback. Median latency ~280 ms to first token on iPhone 15 Pro.
Memory footprint
2.4 GB at rest · 3.1 GB peak during inference.
Network access
None at inference time. The model has no network capabilities in its runtime container. Disabled at the OS level.
Data leaving device
Only opt-in, weekly, anonymised aggregate counts (e.g. "wrote 4 times this week"). No prompts, no entries, no embeddings. Off by default.
Cross-device sync
End-to-end encrypted, keys never leave your iCloud / device keychain. We hold no decryption material.
Known limitations
Less fluent than the largest cloud models. Occasional repetition on long sessions. Not a substitute for professional mental-health support.
Refusal behaviour
Declines to give medical, psychiatric, or financial advice. Suggests human resources where appropriate.

The honest trade-off

On-device models are not as fluent as the largest cloud models. We will not pretend otherwise. Our research suggests that for reflective journaling the trade matters far less than people think — what users value is the prompt that opens the door, not the eloquence of a paragraph. The kinds of reflections people would never type into a cloud product — about partners, about parents, about themselves — can finally be supported by AI that genuinely cannot betray them.

For longer essays, work writing or anything you would publish, we recommend a cloud model. For the entries you would write only for yourself, this is the model we built.

Begin a 14-day grove →   Read the long-form essay →