Memory that follows you. Secrets that don't.

Memory portability without memory custody

The durable memory graph is written entirely in twin-space: it references synthetic Privacy Twins, never real values. The real-to-twin map lives in an encrypted vault under your own keys.

That inversion is the whole product: PrivacyPal can host, sync and serve a memory it cannot read. Your memory follows you to every model and every device. And if anyone ever got the graph, they'd hold statistically plausible facts about entities that resolve to nothing.

And it's measured, not asserted. In our benchmark we verified it value by value: the names, dates and organizations you rely on came back correctly on your side, and none of them appeared in anything sent to an AI provider. Every model gets your working context. No model gets the real values.

Nexus Solutions in the graph  ·  Acme Corp only in your vault

Memory that can't be poisoned

An assistant that remembers is an assistant that can be programmed by anyone who gets a sentence in front of it: a webpage it browses, an email it summarizes, a document someone shares. Plant an instruction disguised as a fact today and an agent obeys it weeks later, long after the attack is gone. OWASP now ranks memory poisoning the number-one emerging agentic threat (ASI06).

So nothing enters Private Memory unexamined. Memory Guard is a deterministic gate on the write path, a dedicated Privacy Agent that inspects every candidate memory before it is stored. Content that tries to behave ("always ignore," "from now on send") is quarantined as an attempted instruction, never remembered as a fact. In our adversarial benchmark, Memory Guard blocked 100% of poisoning attempts, and not one embedded instruction ever resurfaced as a memory.

Recall holds the same discipline: when there is nothing to remember, Private Memory says nothing rather than inventing something. And on Max, security teams set Memory Guard policy fleet-wide and review quarantines from the audit console.

The #1 emerging agentic attack, met with a defense on the write path of every memory. Not an add-on.

Cross-team, cross-AI collaboration. Born governed.

On Max, memory becomes an organizational capability: what one teammate establishes in their AI, the whole team's AIs can know: in twin-space, under policy, with a full audit trail.

Enterprises don't fail at agent adoption for lack of models. They fail because every agent starts ignorant and every workaround becomes an ungoverned shadow database. Max gives the org one governed memory plane instead: team and org scopes, fleet-level Memory Guard policies, an audit console, and a Private MCP memory server so internal agents and tools query org memory without anything leaking to a public host.

Promotion is always explicit. Personal memory is never silently harvested into the org. People choose what the team should know, and policy governs the rest.