HA5H gives your AI tools persistent memory of client matters from a single local SQLite file — SimHash + FTS5 retrieval, no cloud vector database, no embeddings API, no BAA with a third-party store. Privilege stays on your infrastructure, and you can prove it in the room.
The entire matter memory is one SQLite file. Copy it, encrypt it at rest, shred it on matter close. No vector DB to stand up, secure, or breach.
Retrieval runs in-process: zero network calls. The on-prem build fuses outbound connections off — the privacy claim is testable, not a promise.
Every recall shows its per-facet score (SimHash · keyword · salience). Defensible and auditable — no black-box embedding you can't explain to a partner.
Drop the open-source HA5H library into your own environment. One SQLite file per client or per workspace. MIT licensed, no per-seat fees, no vendor in the data path.
The same engine with egress fused off, for privileged work that legally cannot leave the device. What we use to validate the privacy guarantee — and what you can demo to your security team.
# one file per client matter — no infra pip install ha5h from ha5h import Crystal m = Crystal.open("matters/acme.ha5h") m.crystallize("PRIVILEGED — IPR petition due 2026-07-20", salience=5) m.recall("when is the IPR deadline") # local, <1ms, no network
Demo content is fictional and synthetic — no real client data. 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.