The flow, end to end
AI learns. Memory is explicit. Classification is automatic. Policy is deterministic. And only the right memory reaches the right person.
AI
ChatGPT, Claude, Cursor or a custom agent calls remember.
Memory
An explicit, raw memory is submitted — nothing scraped.
Automatically classified
Scope, sensitivity, domain and tags inferred by a cheap model.
Policy enforced
A deterministic engine — never the model — decides access.
Shared organizational memory
One governed store across every AI your team uses.
Right memory → right person
Recall returns only what the human principal may read.
Write
A human, through an agent, sends an explicit memory: content plus optional scope hints and an idempotency key.
remember({
content: "Project Falcon switched to Supplier B.",
scope: "project",
idempotency_key: "falcon-2026-10-04"
})Classify
A cheap decision model proposes sensitivity, domain, tags and importance. Results are stored with provider, model, version, spec hash and probabilities.
{
"sensitivity": "internal",
"domain": "Engineering",
"confidence": 0.94,
"needs_review": false
}Govern
Org, team and project policy compile into an AccessPolicy that may only tighten the classification. Governance is deterministic and testable.
sensitivity_floor: internal
domain_guards:
hr: { min_sensitivity: confidential }Recall
Retrieval builds candidates, then authorizes each one with evaluate(). Only survivors are reranked and returned — with an explanation for the human.
evaluate(principal, memory) -> {
allowed: true,
rule: "project_membership",
summary: "Member of Project Falcon"
}