What you're looking at
Each blob is a topic cluster dominant in that month's conversation history —
extracted from 21,500 documents in Mnemos
using a Bayesian Gaussian Mixture Model per period, embedded with
all-MiniLM-L6-v2. Blob size encodes weight × log(doc_count).
Colour encodes category: green for project/research, amber for engineering, violet for personal, grey for general.
Press ▶ or use ← → arrow keys to step through months. Hit Sound to enable the audio layer — each blob entry fires a pitched tone, and the ambient drone scales with memory density. Empty months whisper rather than silence.
The corpus
| Metric | Value |
|---|---|
| Total documents | ~21,500 |
| Periods | 2023-02 → 2026-03 |
| Peak density | Jun 2025 · 2,952 docs |
| Empty periods | 13 of 38 months |
| Topic model | BayesianGMM · 30 components |
| Embedding model | all-MiniLM-L6-v2 |
Sound design
The Web Audio API synthesises distinct sounds per category and event:
blob arrivals fire pitched tones (F5 for project, C5 for engineering, E♭5 for personal, G4 for general),
staggered so dense periods like June 2025 sound like a chord rather than noise.
The ambient drone at 55Hz / 56.2Hz produces a 1.2Hz beating pattern; gain scales
logarithmically with doc_count so dense periods have real acoustic weight.
Empty periods produce a bandpass-filtered breath — not silence, but hollowness.
No audio libraries. No dependencies beyond D3 and the browser.