OpenMemory
A place for the things you don’t want to forget. Natural notes become structured memories, with you making the final call.
Inside the app · fictional demonstration memoriesA complete path from note to memory.
Text or optional ElevenLabs cloud transcription → classification → separate extraction → human review → an explicit save to a browser vault, with optional encryption.
The hosted app starts with manual capture and explicit OpenAI opt-in. Local inference uses Ollama or a paired, running Mac companion for native MLX.
Fine-tuning, with the evidence attached.
I directed an AI-assisted build and ran MLX LoRA training on a frozen, 4-bit Qwen3.5-9B base: 730 examples and 638,976 trainable adapter values.
| Comparison | Joint accuracy |
|---|---|
| Unadapted base | 67.0% |
| Fine-tuned adapter | 76.5% |
| Constant majority baseline | 81.0% |
A 9.5-point gain over its base; below the majority baseline. AI-assigned labels, class imbalance and regressions limit the result. Experimental, not production-qualified; the released adapter does not classify TV series.
Watch the 91-second walkthrough
Female-narrated, edited captures of the real app. Fictional book and movie memories; not a latency benchmark. Read the transcript ↗