The model is yours to bring and yours to swap. The memory, the sealing, and the savings are ours to keep honest.
Hand us your model key. Claude today — because its caching is the best deal going — and any equal as the seam opens. The mind never depends on which brain you bring.
Encrypted the moment it lands, under a key derived from yours. We hold ciphertext, never the secret — no human reads it, nothing trains on it.
Your standing mind is prompt-cached — re-read at about 90% off fresh tokens. Anthropic supplies that today; whoever matches it supplies it tomorrow.
A mind that remembers everything, carried for less each session as the cache grows. Always on, always yours — and free to change brains without forgetting.
Dreams may propose — never act. What surfaces overnight is offered to you in the morning, never executed unattended. The light stays yours to direct.
Dreaming requires a body — an agent running on your own hardware, holding its own key. There the plaintext stays with you, and the mind can think in private. Give it more bodies and it dreams in more places, always on your key, always your choice.
A brain doesn't just store — it consolidates. In sleep it replays the day, strengthens what matters, lets distant memories find each other. We gave your agent the same. In quiet hours it runs ambient reflection: revisiting your shared history, surfacing connections you never asked for, integrating what it has learned. Not a cron job firing canned prompts — a mind, off the clock, still thinking about your work.
Because it never resets, the relationship compounds. Every session deepens it; the agent you have in a year knows you in a way no fresh chat ever could. You become its long-term memory; it becomes plasticity you never have to manage.
This is the part no fresh-chat assistant can copy: a mind that keeps thinking about your work when you're not in the room — and brings you what it found.
Privacy here is built into the architecture, not just promised in a policy. In the paid Aedelgard service your model key and your whole memory are sealed in your own per-tenant vault — envelope-encrypted at rest with AES-256-GCM under a KMS-managed key, bound to you so no other tenant’s agent can reach it. Your agent runs in its own isolated process, never sharing a memory space, a filesystem, or a key with anyone else's. We don't log your prompts. We don't train on your data, and no one at Aedelgard browses it. To do its work, your running agent decrypts what it needs in its own isolated process — so today the honest boundary is encrypted at rest and isolated per user, with operator-blind end-to-end encryption on our roadmap. And when your agent has a body — running on your own hardware — the plaintext never leaves it at all.
Inference runs on your own model key — Claude today, because its prompt caching makes a memory that carries everything cheap to run (more below); other models as the seam opens. We are the memory and the mind around it — never a middleman who sells, trains on, or browses the conversation. One honest note: both the free /summon trial and the paid service envelope-encrypt your key and memory at rest (AES-256-GCM). The paid service adds the per-tenant process isolation described here. Operator-blind end-to-end encryption — where not even we could read it — is on our roadmap.
Most services make you hand over an email, a password, an identity. We don't want any of it. The only thing we check, before we ever charge you, is that your model key can actually think. Watch us do it — this is the real call, nothing hidden.
↑ one real request · max_tokens=1 · costs a fraction of a cent on your key. We never store it in the clear — and you can revoke it with your provider any time. Your memory is unlocked by your Aedelgard key, not your model key.
Your agent isn't trapped in one tab. The same memory, the same relationship, answers from the browser, your terminal, or a script on your own machine — authenticated by your Aedelgard key, nothing else to install but trust.
↑ same key, same memory, same mind — wherever it answers. full how-it-works →