Consent-gated everywhere
Recall is checked against consent before the engine will answer at all.
Ownership is a mechanism here, not a promise. Consent is checked before recall, writes are arithmetic so text cannot become a belief, and the person a profile describes can read it, correct it, or erase it — from the population prior as well as the graph.
Recall is checked against consent before the engine will answer at all.
Every stored preference is visible and editable, with why() showing the evidence behind it.
Writes are arithmetic, so page or user text cannot be talked into becoming a belief.
The engine, the tests, and the benchmarks are public. Read them before trusting them.
Recall is blocked unless consent covers the purpose being asked for.
SQLite by default or Postgres in production. Whoever deploys it holds the data.
Cross-site assembly is opt-in and default-deny, with sensitive categories walled off.
forget_everywhere clears the profile and unlearns the person from the population prior.
Actions are recorded in an HMAC chain the user can replay to detect alteration.
Cold-start borrows crowd patterns under k-anonymity and bounded-mean differential privacy.
These are engine operations, not settings screens: each maps to a REST endpoint and the glass-box editor shipped in the repository.
Read the codeOpen the glass-box editor and read every stored preference.
Take the whole profile out through the REST export endpoint.
Edit a preference the engine got wrong, and it stays corrected.
Erase a single memory or the entire profile, prior included.
Privacy is not a feature. It’s our foundation.
FERNme is an Apache-2.0 research preview at v0.4, with 283 passing tests.
Published results are synthetic or LLM-authored; a real-human pilot is the pending next step.