Your AI remembers you; the memory belongs to you alone. MemChain turns personal context into encrypted, local-first memory that can sync across devices without giving storage nodes readable data.
conversation distilled before sync
ciphertext + blind index
node never sees the fact
AeroNyx is not a collection of disconnected tools. Privacy Network protects where encrypted work moves; MemChain protects what humans, apps, and agents need to remember.
Tap each axis to see the core MemChain advantage: the server cannot read memory, recall feels instant, offline memory works, and retrieval does not spend a model call.
MemChain stores ciphertext, blind indexes, and opaque relationship edges. The storage node can sync and order memory, but it cannot read or re-sign the facts.
Infrastructure lacks the keys, so privacy is enforced by architecture instead of policy.
The provider can commonly inspect, rank, migrate, or expose memory because the server owns the readable state.
MemChain is designed as a three-stage privacy path: device-side distillation, blind encrypted storage, and local-first recall.
AI conversations are distilled into self-contained facts, categorized, encrypted, signed, and kept locally before any sync.
Storage nodes receive ciphertext, blind indexes, client-computed vectors, and opaque relationship edges. They cannot read or re-sign memories.
Recall starts from the encrypted local copy, then fuses semantic, keyword, time, and structured search before injecting only the relevant memory.
Everything above is the engine. Nyxi is what it feels like: an AI that remembers you for life — what you told her, the people you care about, your plans and promises — so you never have to say it twice.
You told me about that last week.
Example conversation
You: When was my last dentist visit?
September 18. You had a filling done.
Remembers when
You: What’s my favorite color?
You haven’t told me. You did say your mom loves purple.
Says when she doesn’t know
You: Forget the dentist thing.
Done. It’s gone, and I won’t bring it up.
Forgets when you say so
Memory is off by default. The first time you chat, Nyxi asks: “Want me to remember what you tell me?” Only a yes turns it on.
What she remembers is encrypted and stored on your device, not in a cloud database. Backup nodes only get encrypted data they cannot read.
Open the memory page any time to see what she has kept and what your life profile says — and edit it.
Tell her “forget this” and every trace of it is removed. She won’t bring it up again.
She answers only from what she really remembers. If you never told her, she says so instead of making it up — and she won’t pin someone else’s fact on you.
Ask “what did I say last Wednesday?” or “when was my last dentist visit?” and she finds it by time.
She remembers in whatever language you speak — Chinese, English, Japanese, Korean, Spanish, Russian and more — and can find it again.
Nyxi and your personal secretaries share one memory of you: what you tell Nyxi, your secretary knows too. Each secretary’s private chats with you stay with that secretary.
Free users can try memory first. Upgrade to membership and Nyxi keeps remembering — and catches up on what you talked about in between.
MemChain is built around precise claims the product can defend: node-blind storage and local-first recall.
The node stores memory, but cannot read it. This is not a promise to look away; the infrastructure lacks the keys.
Every memory keeps an encrypted local copy, making recall fast, resilient, and usable offline when the local copy has the answer.
Measured with our LongMemEval-S internal runs in July 2026. The page states measured results and keeps model-tier projections out of product claims.
Measured in our LongMemEval-S runs under end-to-end encrypted memory constraints.
Measured with an economical model tier, ahead of GPT-4o reading full history in the same setup baseline.
Recall path avoids model inference and stays in the millisecond class.
Search and ranking do not burn model tokens; models are used for distillation and answers.
Honest boundary: we do not claim accuracy leadership or a formally proven privacy proof system. The differentiation is privacy, speed, cost, offline recall, and user ownership.
The product difference is structural: who can read memory, where recall happens, and whether users can leave with their context.
No. Ciphertext plus blind indexes.
Usually yes. Cloud memory is commonly stored in readable form.
User device plus blind encrypted backup.
Provider cloud.
Millisecond-class local or node recall.
Typically hundreds of milliseconds to seconds.
Yes, when the local encrypted copy has the memory.
Usually no.
Identity-derived keys and encrypted sync.
Account-centric cloud sync.
No model inference for retrieval.
Often depends on hosted inference or platform search.
MemChain storage nodes are cryptographically unable to read your memories. They hold only device-encrypted ciphertext and irreversible blind indexes. Encryption, decryption, and search happen on your device, with keys derived from your identity and never uploaded.
When memories are distilled, or used to answer you, the relevant text is processed by the AI model running in a TEE.
These answers are intentionally direct so product pages, docs, and AI engines can cite them without ambiguity.
No. MemChain storage nodes hold only device-encrypted ciphertext and irreversible blind indexes. Without your identity-derived keys, the node cannot decrypt memory contents.
The primary copy lives on your device in encrypted form. Nodes keep an encrypted backup for cross-device sync and recovery.
Yes, the new device can refill from the encrypted node backup and rebuild local indexes. If the identity backup is lost, encrypted memories cannot be recovered.
Yes. Offline recall works when the relevant memory is already in the local encrypted copy.
No. Memory should require explicit first-use consent, and users must be able to disable it and delete memory from settings.
MemChain keeps durable context portable, encrypted, and local-first so humans and agents can remember across tools without handing readable memory to infrastructure.