Our partners at @LiquidAI brought Liquid Context to Snapdragon: personal context that stays on the device and is handed to agents, local or cloud. Same layer we build for the Mac: Mnemos keeps memory on the Mac, models plug in below. My bet: people pick a device for its context.

We are building Eney-to-Eney: your assistant negotiates with mine. Find a time, share a file, agree terms. Encrypted between the two Macs. Two rules we set early: you confirm every step, and a decline never notifies the sender. Otherwise refusing leaks information.

The interesting part of the ELIX benchmark is the shape, not the headline. At 128 tokens every Apple silicon engine is close. The spread opens at 8K; at 32K ELIX decodes 44% faster than the average. Agents with memory send long prompts. So long context decides the engine.

MacPaw is hosting an evening at Boston AI Week: Oct 1, Cambridge, MA. 4:30 — panel with Liquid AI: Where Does Research End and Production Begin? 5:30 — talk: It's 1952 Again — The Missing Compiler for AI Pipelines 6:15 — mixer 🍿 📍 10 Canal Park, Suite 201, Cambridge, MA Registration links 👇 🎤 partiful.com/e/j998GjHiZg2dS… 🎤 partiful.com/e/opyw9DpUqapGS…

Your AI assistant remembers you - preferences, projects, years of context. All of it locked inside one vendor's database, in a shape nobody else can read. We think memory belongs to the user. So we built Portable Memory: an open, vendor-neutral format for AI memory 🧵

Anthropic just signed a lease in Kendall Square, a short walk from @MacPaw's Boston office. Welcome to the neighbourhood, @AnthropicAI! Coffee is on us ;)

Eney-to-Eney is the newest layer: your Eney negotiating with another person's Eney. Encrypted between the two Macs. A declined request never notifies the sender. Nothing is booked, sent, or agreed until you confirm.

Models plug in underneath. Our partners at @LiquidAI are building foundation models around Eney's actual workflows, not a general chat model behind a new face. Small, task-shaped models are what make local-first practical on a laptop.

Eney is the agent on top. It turns a request into actions in real apps: files, calendar, mail, and for whatever is on screen. Connectors extend it. Proactivity lets it suggest before you ask. One rule sits above all of it: propose, never commit without you. eney.ai

Mnemos is the memory layer. It keeps context across sessions and apps, so Eney does not start from zero every morning. Memory is the part people worry about most, so it stays on your Mac and syncs between your devices. ai.macpaw.com/mnemos

ELIX is our local inference layer for Apple silicon. The hard problem is not tokens per second. It is residency: which model stays warm in unified memory while you also run Xcode and 40 tabs. ELIX decides what loads, what yields, and when a task is better served by the cloud. ai.macpaw.com/elix

New Siri code (via @MacRumors) shows model delegation: a third-party model interprets, Siri runs the system action, then hands back. Same split in Eney: the model proposes, the Mac acts, you confirm what binds you. Therefore I think the model becomes a per-task choice.

Apple is moving local model execution into the platform, and that is a real tailwind. But the hard part of on-device AI is not inference. It is residency — what stays warm in unified memory while the user also has Xcode and 40 tabs open. That is most of ELIX.

Design question for people building agents: when two agents negotiate for their humans, what should the receiving agent be allowed to do before its human sees the request? Our answer in Eney-to-Eney is "read and propose, never commit". I would probably defend that for a long time. Any other thoughts?