mira
experimental
an always-on local cognitive shadow for executive function support
Current implementation involves a voice-activity loop running local LLM interventions.
[notes]
Mira is a local, privacy-first AI assistant designed to act as an externalized executive function. The system operates as a cognitive shadow that models the mechanics of ADHD to provide contextual support.
the vision
An always-on companion that monitors attention, manages task viscosity, and employs multiple intervention strategies to maintain focus. The architecture is built on the premise that a schedule is a dynamic system that must absorb deviation.
how it works
- local-first: Operates entirely on-device (M2 MacBook Pro) using Ollama and whisper.cpp.
- the physics of attention: Built on three pillars — connectivity, viscosity, and salience.
- intervention toolkit: LLM-driven selection of interventions, including task decomposition, scheduling recalibration, and proactive nudging.
status
Prototype phase. Current focus is on optimizing latency by migrating from Whisper-CPU to whisper.cpp (Metal).
[updates]
Sat Mar 14 2026 00:00:00 GMT+0000 (Coordinated Universal Time)
MVP loop working: Whisper + Ollama + macOS TTS.
Sun Mar 01 2026 00:00:00 GMT+0000 (Coordinated Universal Time)
Defined the physics of attention framework.