Private Memory: memory that follows you across every AI

The problem

Four assistants. Four strangers.

ChatGPT remembers inside ChatGPT. Gemini builds its own profile. Copilot remembers inside the mailbox. Switch models and you start from zero, so your best context stays trapped wherever you happened to type it. Nobody has solved memory across the gardens, because whoever holds cross-provider memory holds the most sensitive longitudinal dataset in your working life. No model vendor is structurally trustable with it. PrivacyPal is built to be exactly that owner, because our memory never contains a real secret at all.

Today: memory in walled gardens

ChatGPTknows Monday's brief

Claudeknows Tuesday's draft

Geminiknows one meeting

Copilotknows your inbox

Four partial pictures. Zero portability. Re-pasting context is how data leaks.

With Private Memory: one memory, every AI

YOUR MEMORY · twin-space · your keys

ChatGPTfull context

Claudefull context

Geminifull context

Copilotfull context

One canonical memory, injected into whichever AI you open, plus Claude Code and your own agents.

The mechanism

Captured where you work. Stored where nobody can read it.

Private Memory runs on the interception plane PrivacyPal already operates, the same place Privacy Twins swap your secrets out of every prompt.

Capture

Every AI exchange (prompt and response, whichever model served it) is captured at the interception point with zero added latency in the hot path.

Consolidate

A background "sleep-time" process distills durable facts and reconciles them: new facts supersede old ones, contradictions are invalidated, never silently deleted.

Recall

At the moment you prompt any AI, relevant memory is assembled into a token-budgeted Memory Pack (no LLM in the read path) and injected in twin-space.

Erase

Destroy your vault keys and every copy of your memory (synced, exported, cached) becomes meaningless noise. The right to be forgotten, as mathematics.

01 · The trust model

Twin-space storage

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.

Zero real secrets in the graph Vault keys held by you Bi-temporal: facts supersede, never vanish Crypto-shred erasure

Memory Pack · injected at prompt time · twin-space

· Active deal: Nexus Solutions renewal, decision expected Q4

· Contact: Dana Brooks (CFO) prefers numbers in tables

· Your style: concise, mechanism-first, no adjectives

· Project Kepler: budget approved, kickoff pending

// twins resolve to real values only on your device

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

Real secrets in the memory graph. Facts reference twins; reals stay in your encrypted vault.

+0ms

Added to the hot path. Capture and consolidation happen off the request path, in the background.

LLMs in the recall path. Memory Packs are assembled by deterministic retrieval: fast, and nothing to leak.

Action to erase everything, everywhere. Crypto-shredding makes deletion a property of math, not a promise.

For individuals · PrivacyPal Pro

Your canonical context, in every AI you open

Stop re-introducing yourself to machines. With Pro, your projects, preferences and working context become one portable memory: on-device, opt-in, and yours.

Continuity

Start in ChatGPT. Finish in Claude.

Brief one model on Monday and every model knows it Thursday. The project, the constraints, the decisions so far: your context is simply there.

Monday you tell ChatGPT about the Kepler kickoff and its budget ceiling. Thursday you open Claude and say "draft the kickoff agenda". No re-pasting, no re-explaining. It knows.

Freedom

Pick the best model per task

Memory is how vendors lock you in. When your context travels with you, you choose models on merit (reasoning here, drafting there) without amnesia as the switching cost.

Use Gemini for research, Claude for writing, Copilot in the inbox: all four working from the same canonical memory of who you are and what you're building.

Control

See it. Steer it. Shred it.

A full viewer shows everything remembered and why. Go incognito before a sensitive session. And when you say forget, one action erases every copy, permanently.

Open the memory viewer to browse and prune facts, flip on incognito for a session that should never be remembered, or crypto-shred the lot in one click.

02 · The enterprise

PrivacyPal Max

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.

Explore PrivacyPal Max See it on your org's workflows

Personal scope

An AE's discovery call insights live in their own memory, captured from their Claude session, private by default.

↓ explicit promotion · policy-checked · audited

Team scope

Promoted to the deal team: the SE's Copilot and the CSM's Gemini now answer with the same account context: no forwarding threads, no re-briefing.

↓ org policy plane · Memory Guard · audit console

Organization scope

Institutional knowledge (terminology, standards, hard-won process) inherited by every governed agent from day one. New hires' AIs start informed, not ignorant.

Memory Guard

Memory that can't be poisoned

Persistent memory is the newest attack surface in agentic AI: plant an instruction disguised as a fact today, and an agent obeys it next month. The OWASP agentic security taxonomy ranks memory poisoning among its most critical risks.

Deterministic write gate

Every candidate memory passes Memory Guard before it's written: a dedicated Privacy Agent on the write path.

Instructions ≠ facts

Content that tries to behave ("always ignore," "from now on send") is quarantined, not remembered.

Fleet policies on Max

Security teams set org-wide Memory Guard policy and review quarantines from the audit console.

Where you get it

Private Memory across the ecosystem

Pro

Personal cross-LLM memory, on-device. Viewer, incognito and one-action erase included. Opt-in, off by default.

PrivacyPal Pro

Max

Everything in Pro plus team & org scopes, the policy plane, fleet Memory Guard, Private MCP memory server and audit console. Rolling out now.