Donna

The reusable idea

Separate the capture layer from the answer surface. Most assistants make capture loud (notifications, check-ins, integrations) and answer cheap (a single query). Donna inverts it: capture is silent and continuous, the answer is the only thing you experience. The product is the gap between "how does she know that?" and the fact that she does.

What it is

Donna is a local-first macOS ambient assistant. It watches your screen and your DMs, builds a model of your work over time, and answers when you ask. The design choice that defines it: reactive, not voice-first, not proactive. No nudges, no notifications, no wake word. Screen capture, OCR, redaction, and memory accumulation all run silently. You only ever touch the answer surface by DMing a Discord bot. State stays under ~/.donna/ and nothing leaves the machine except the calls the enabled mode actually needs.

The engineering that actually takes

The hard problem is turning a raw screen feed into durable, retrievable memory without leaking secrets and without paying for a vision-LLM call on every frame. Screenshots are noisy, redundant, and full of API keys. The pipeline is: deduplication, Apple Vision OCR, then a pure redaction pass before anything is stored.

Redaction is two stages. First: known key prefixes (sk-ant- before sk-, so Anthropic keys are not swallowed by a generic OpenAI pattern). Second: a Shannon-entropy heuristic for unknown secret shapes, with hex and UUID carve-outs to cut false positives.

Projects are not configured; they are discovered. donna screen summarize embeds transcripts with a local Ollama model and runs single-linkage cosine clustering in pure numpy. The algorithm is deterministic and order-independent. An autonomous distiller then asks the model what is worth remembering long-term, clamps a confidence score, and writes it to SQLite with FTS5. The distiller swallows every per-row failure so one bad summary never stops the run.

Honest limitations

The brain still calls Gemini, so Donna is local-first but not local-only. Entropy-based redaction is a heuristic: it will miss low-entropy secrets (short API keys, short passwords) and over-redact some random-looking strings. Clustering is O(n squared) on the similarity matrix, which is fine at personal scale but not at archive scale. macOS only; the Apple Vision dependency is not portable.