On-device · no PHI egress · no BAA needed

Patient memory that never leaves the device.

HA5H gives clinical AI tools persistent memory of patient charts from a single local SQLite file — SimHash + FTS5 retrieval, no cloud vector database, no embeddings API. PHI stays inside your four walls, so there's no third-party data processor to add to a BAA — and you can prove it.

Live demo · runs in your browser
Ask a patient chart. Watch where the PHI goes — nowhere.

One file, not a stack

The whole chart memory is one SQLite file on the device. No vector DB to stand up inside your compliance boundary — or breach.

No PHI egress, provably

Retrieval runs in-process: zero network calls. The on-prem build fuses outbound connections off, so chart text physically cannot reach an external API.

Transparent ranking

Every recall shows its per-facet score (SimHash · keyword · salience) — auditable retrieval, not an opaque embedding service.

How it deploys

Self-host the library — or run it air-gapped for data that can't touch a network.

Recommended · self-host

Drop the open-source HA5H library inside your existing compliance boundary. One SQLite file per patient or per workspace. MIT licensed, no per-seat fees, no vendor in the PHI path.

Strict mode · air-gapped

The same engine with egress fused off, for PHI that legally cannot leave the device. What we use to validate the privacy guarantee — and what your security team can test.

# one file per patient — PHI never leaves the box
pip install ha5h
from ha5h import Crystal
m = Crystal.open("charts/mrn-0942231.ha5h")
m.crystallize("metformin 500 mg BID; penicillin allergy", salience=5)
m.recall("current metformin dose")     # local, <1ms, no network

Demo content is fictional and synthetic — no real patient data or PHI. The in-browser demo shows HA5H's actual ranker output (curated queries use exact precomputed scores; free-typed queries are scored client-side over the same file) and makes no network calls after load. HA5H is software, not legal advice; your HIPAA compliance depends on your overall deployment.