HA5H gives financial AI tools persistent memory of client accounts and live deals from a single local SQLite file — SimHash + FTS5 retrieval, no cloud vector database, no embeddings API. Client PII and MNPI stay inside your information barrier, with no third-party processor in the data path — and you can prove it.
The whole book of memory is one SQLite file — copy, encrypt, retain (17a-4-friendly), shred. No vector DB inside your compliance perimeter to secure or breach.
Retrieval runs in-process: zero network calls. The on-prem build fuses outbound connections off, so MNPI and client PII physically cannot reach an external API.
Every recall shows its per-facet score (SimHash · keyword · salience) — auditable retrieval for compliance, not an opaque embedding service.
Drop the open-source HA5H library inside your existing perimeter. One SQLite file per client, deal, or desk. MIT licensed, no per-seat fees, no vendor in the PII/MNPI path.
The same engine with egress fused off, for restricted/MNPI work that legally cannot leave the device. What we use to validate the guarantee — and what your compliance team can test.
# one file per client/deal — nothing crosses the barrier pip install ha5h from ha5h import Crystal m = Crystal.open("desks/ma-advisory.ha5h") m.crystallize("MNPI — RESTRICTED: do not trade CRVH/HLSM", salience=5) m.recall("which names are on the restricted list") # local, <1ms, no network
Demo content is fictional and synthetic — no real client data or MNPI. 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 compliance advice; your regulatory posture depends on your overall deployment.