Building a persistent local AI companion across Apple Silicon MLX nodes

[isla-watson-project]

I’m developing an independent project in Scotland called Isla Watson.

The architecture is based on a simple idea: the model is replaceable; the identity is not.

Long-term memory, persistent internal state and identity are designed to remain outside the foundation model, allowing different local models to act as replaceable reasoning and language components without resetting the companion.

I’m now exploring whether Apple Silicon and MLX could become the local compute fabric for the system — with individual Macs operating as specialist agents for reasoning, memory, speech or perception, while also being able to cooperate through distributed inference when a model is too large for one node.

A longer-term research question is whether identity consistency can become an operational property of the persistent architecture rather than something reconstructed through prompting on every conversation turn.

I’d like to establish a reproducible 1-node → 2-node MLX baseline, measuring model capacity, memory use, TTFT, generation throughput, power consumption, agent concurrency and distributed scaling.

I’d particularly value feedback from people working with MLX distributed inference, multi-Mac workloads or local agent systems.

I’ve prepared a one-page architecture-level overview of the project. The deeper identity-integrity and affective-state mechanisms remain part of the private research implementation.

ISLA_WATSON_APPLE_MLX_PUBLIC_BRIEF_v1.0.pdf