Why can't my phone's AI assistant actually remember who I am?
The core architectural decision was to keep the ZeroClaw memory file
format exactly as-is. soul.md, memory.md, agents.md — same plain text
format, same paths, just stored in Application Support/cue/memory/
instead of ~/.zeroclaw/. This meant the iOS app would be
compatible with the desktop agent's memory files. You could theoretically
sync them with iCloud Drive and have continuity across both.
I used GRDB.swift for the SQLite layer. It gives you a Swift-native
interface and WAL mode out of the box for concurrent reads, which matters
when the agent loop is running on one thread and the UI is reading memory
on another. All the API keys go into Keychain with
kSecAttrAccessibleWhenUnlockedThisDeviceOnly.
The ZeroClaw agent loop is written in Node.js with an async event loop managing tool calls. Swift's actor model is actually a great fit for this — AgentActor is a Swift actor that handles the entire turn loop serially, with async/await for the inference calls and AsyncStream for streaming responses back to the UI.
Tool calling works via structured output: the model generates a
ToolCallEnvelope JSON struct, and CueDispatcher routes
it to the matching skill. The skill executes, returns a result string,
and the loop re-prompts with the result. The hardest part wasn't the
logic — it was getting Apple Foundation Models to produce reliable
structured output for tool calls. Guided generation helped a lot.
iOS has a stricter permission model than desktop, which shaped how the skill system works. On desktop, ZeroClaw has a shell execution skill. On iOS, there's no equivalent — the sandbox won't allow it. Instead Cue uses Apple Shortcuts via AppIntents as the OS automation layer.
The bundled skills cover the things I actually want an assistant to do
on my phone: add calendar events, set reminders, search the web, read
and write files via the document picker, and manage the clipboard.
Each skill is a CueSkill protocol conformance, registered
with SkillRegistry on launch.
Cue has been one of the most satisfying projects I've worked on because it's solving a problem I actually have. Porting the ZeroClaw runtime to Swift taught me a lot about how Swift's concurrency model works in practice, and working with Apple's Foundation Models API gave me a much better sense of what on-device AI can actually do today.
The serialized execution model of Swift actors maps almost perfectly to what you want in an agent loop — one turn at a time, no concurrent mutations to state, async/await for I/O. The final architecture is clean in a way that the Node.js version wasn't.
The right language feature for a problem makes the code almost explain itself.
Deciding early to keep the ZeroClaw memory file format exact was the right call. It unlocked cross-device compatibility for free and made the iOS app feel like a genuine extension of the desktop experience rather than a separate product.
Compatibility is a feature — don't break formats without a good reason.
Apple Foundation Models on iPhone 15 Pro are genuinely capable for everyday assistant tasks. For most things users actually want — summarize this, create a reminder, what was I working on — on-device inference is fast, private, and good enough.
The infrastructure matters more than the model size for most use cases.
