A useful private agent needs durable context, but sending an entire personal model to a backend defeats the premise. The problem is selective recall: enough relevant memory to help, no ambient plaintext collection, and an honest erase path.
Select a minimal, query-relevant memory card locally; decrypt it only for the owner’s current session; and attach enough provenance for an answer to explain its grounding without exporting a private corpus.
Retrieval quality improves when a system has more context. Privacy usually improves when it has less. Treating that as a budgeted systems problem makes the trade-off explicit.
The web vault holds its AES-256-GCM key and owner token in memory only and clears them on identity change, expiry, or lock. Portable, query-selective memory retrieval is active work, not a complete product claim.
Primary source: Personal World Model research
Prototype a local relevance gate with a fixed context budget, then measure answer quality, leakage surface, and failure behavior against a deliberately over-broad baseline.
The companion essay is written for a broader technical audience. The repository and developer community are the places to turn a claim into a contribution.
🤫 One is made by Hushh Technologies Corporation, an independent company. We name the hardware and clouds One runs on to say where it runs. None of them endorse us, and we call a company a partner only once the agreement is signed.