How it works
Four steps before anything lasting changes.
Gromani is not a vector database. It is the decision layer in front of whatever store and model you already use.
01
Ingest
The agent proposes a memory: a statement, its subject, the channel it came through and who said it.
02
Classify
Gromani attaches a confidence score, a privacy class and the conditions the memory applies under, and looks for an existing memory on the same subject.
03
Decide
A policy chooses retain, update, forget or escalate. Authority, scope and consequence decide whether a change is lasting, a scoped exception, or a question for a person.
04
Log
Every decision writes an audit entry: what was proposed, what was decided, by which rule, and what it replaced. Removed memories leave a tombstone.
A policy in seven rules
The toy policy behind the demo on the home page. Real policies are per deployment, versioned, and part of the audit trail themselves. Threshold here: 0.70.
- R1A deletion request from the owner always forgets.
- R2An elapsed retention window forgets.
- R3Restricted data without a recorded approval escalates.
- R4Confidence below the policy threshold escalates.
- R5A conflicting statement from equal or higher authority updates.
- R6A conflicting statement from lower authority escalates.
- R7Otherwise, retain with source and class.
What we will measure
Targets are being set in pilot design. No results are published yet.
- False-memory rate against a fixed baseline
- Inappropriate disclosure rate by privacy class
- Share of genuine changes accepted without review
- Review load: how much reaches a person, and how fast it clears
- Response time and cost per decision