The model we're testing

SignalIntentMemoryActionSkill

A voice ask starts the loop: what arrived, what you meant, what Luke knows, then what runs.

See it as

EverydayBundledEmergentRecall

Fig. 2. The interaction loop under study, replayed through four lenses: everyday, bundled, emergent, and recall.

What Luke does today.

Everything in this list works right now. The first release set out to test one claim: that the signals you give a personal AI can become real help.

• Voice: speak naturally; Luke reads the intent.
• Quiet capture: meetings and ambient work, no surveillance.
• Delegated research: hand off a question, get an answer that keeps working.
• Visible memory: a layer you can open, read, and edit.
• Early skills: the work you repeat, saved and reused.

What real use taught us.

Then people used it. What they leaned into, we deepened. What they pulled back from, we made optional.

Leaned in

Short asks, captured meetings, clean handoffs. We made these sharper.

Pulled back

Capture and memory. So what Luke keeps is visible, and yours to steer.

Each release is shaped by the last one.

Trust layer · audit sheet

Built to be honest, not just sound right.

That loop shapes a standard Luke holds to before it earns more agency, the part you can check, not just take on faith.

Agency readiness

more filled, more settled

• Facts about you trace to a source
• Sources stay visible
• It says when it doesn't know
• Capture can't rewrite your ask
• Actions leave a trace
• Longer tasks stay reliable

Won't

• ✕ Invent facts about you
• ✕ Hide the models or sources it used
• ✕ Let passive capture rewrite your ask
• ✕ Claim a task ran when it didn't

Does

• ✓ Say when it doesn't know
• ✓ Show the models and sources it used
• ✓ Keep actions inside boundaries you can see
• ✓ Leave an auditable trace
• ◐ Keep tightening longer-running tasks

Open edge: longer-running tasks and cross-turn reliability.

Fig. 3. The agency-readiness audit: what Luke refuses, what it does instead, and the open edge still being earned.

Where Luke is headed.

A short, honest look at the direction, shaped by what real use is teaching us. None of this changes what works today.

• True compositionality: many models ranked and composed within one task, turn by turn, not a single model carrying it alone.
• Evolving skills: a task you repeat becomes a skill that improves with every run, each change reviewed before it lands.
• Extended runtime: an optional sandbox you switch on yourself, for longer jobs that keep running while your everyday use stays on your Mac.

Get Luke for macOS A personal AI that remembers who you are and works the way you do.