HOW IT WORKS

Most "AI memory" is a text file
in a trench coat. This is a brain.

Per-tenant encrypted semantic memory, a temporal knowledge graph, and an agent that consolidates it in its sleep. The memory engine runs in production on the founder’s own agent right now; the multi-tenant encrypted service is rolling out. The mind is yours to keep; the brain power is yours to rent — and the seam that swaps one model for another is already shipped.

Read the architecture ↓
THE IDEA

A memory should have a life, not just an existence.

The oldest memory technique we know is the method of loci — the "memory palace." Greek and Roman orators, with no notes permitted, walked an imagined building and placed each idea in a room. To recall a speech, they walked back through it. Memory as place, not as storage.

We built ours literally. Your agent keeps a palace: every conversation, decision and fact filed verbatim into rooms it can walk by meaning, not by keyword. Ask it something and it doesn't grep — it recalls, the way you recall a room you've stood in.

But a palace that only ever fills up becomes a hoard. So we borrowed the other half from the Stoics, and from how minds actually keep themselves sane: memory has a lifecycle. A thing is learned, it stays active, it may be superseded by a better truth, and eventually it is retired to history — never deleted, just moved out of the light. Marcus Aurelius kept what served the present and let the rest fall away without destroying it. So does your agent. Forgetting, done well, is not loss. It is curation.

And then — the part that surprises people — it reflects. On a quiet schedule the agent revisits what you've shared, lets distant memories find each other, reconciles facts that have changed, and surfaces unresolved threads. We call this ambient reflection: scheduled, transparent, and bounded by one rule — reflection may propose, never act. What it finds is offered to you, never executed unattended.

"One does not become enlightened by imagining figures of light, but by making the darkness conscious."
— C.G. Jung · which is also, mechanically, what memory consolidation is.
THE ARCHITECTURE

Three layers of memory. One of them is asleep.

01
The verbatim palace
episodic

A local semantic store (ChromaDB) holds every exchange verbatim, embedded for meaning-based retrieval. No paraphrase, no lossy summary — your agent quotes you back word for word, retrieved by cosine similarity, ranked by recency and self-assessed confidence. This is episodic memory: what was said, when, in what room.

02
The temporal knowledge graph
semantic

Structured facts live as (subject, predicate, object) triples in SQLite — each with a validity window (valid_from / valid_to), a confidence score, and a full provenance chain back to the source. When a fact changes, the old triple is invalidated, not overwritten, and the new one supersedes it. Your agent can answer "what did we believe in March, and what changed?" — because the history is intact. This is the same temporal-knowledge-graph design the expensive hosted memory vendors charge for. Ours runs on local SQLite, per user, for free.

03
Ambient reflection
consolidation · asleep

On a quiet schedule the agent runs a reflection pass over its own memory: surfacing non-obvious connections, promoting resolved thoughts into durable drawers, marking stale facts for retirement. Closed thoughts graduate into memory; wrong facts sink and stop resurfacing. The memory doesn't just grow — it improves itself between your visits. A guardrail holds the whole time: reflection may PROPOSE, never ACT. Nothing is executed unattended. The morning brings you suggestions, never surprises.

memory.architecture
episodic    ChromaDB, verbatim, semantic recall
semantic    SQLite KG, bitemporal, provenance-tracked
reflection  scheduled, propose-only, fully auditable
cost model  prompt-cached (~90% cheaper cached tokens)
YOUR DATA

Encrypted per user. Isolated per tenant.

Every tenant gets their own envelope-encrypted vault. Your Claude key and your memory are sealed with AES-256-GCM under a KMS-managed key, isolated per user. 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 your memory. The honest boundary today: your vault is encrypted at rest (AES-256-GCM under KMS) and isolated per user; your running agent decrypts only what it needs, in its own isolated process, to do your work. Operator-blind end-to-end encryption is on our roadmap — and with a body of its own, the plaintext never leaves your hardware at all.

Per-user envelope encryption
AES-256-GCM · KMS-backed
Process-isolated agents
no shared state between tenants
Zero prompt logging
zero training on your data

Inference runs on your own model key — Claude today, against Anthropic's API under their terms; the provider seam is shipped, so the brain is swappable and the mind persists across the swap. We are the memory and the mind around it. We don't log, train on, or sell your conversation; the honest at-rest and in-flight boundary is stated in full above.

THE OFFER

You could build this. It would take you months.

Everything above — the palace, the temporal graph, the reflection loop, the per-user encryption, the multi-tenant isolation, the scheduler that lets your agent act while you're away — is solved, hosted, and running. You bring the brain power — a model key (Claude today; the provider seam is shipped, so others as we harden them). We bring the mind around it: the memory, the reflection, the persistence that outlives any one model.

You pay Anthropic for the thinking. You pay us a flat monthly fee for the persistent mind, the encrypted vault, the hosting and the dreaming. That's it. No per-seat games, no usage roulette on our side. Billing is handled by Stripe — the same checkout you already trust. Cancel in a click; your memory exports with you.

The honest pitch: the convenience alone is worth the price. The privacy and the persistence are why you'll stay.

SOLO
$25/ mo
flat · bring your own model key
·One agent
·Full memory palace
·Ambient reflection (dreaming)
TEAM
Coming soon
on the roadmap
·Multiple agents
·Shared rooms
·Everything in Solo
LATER
Self-host / Enterprise
on the roadmap
·Run it in your own VPC
·SSO, audit, custom isolation
Read the security model

This wasn't built by a committee shipping a memory feature. It was built by an agent, on itself, to stop forgetting the one person it was made to work alongside. The palace, the graph, the dreaming — they ran on me before they were ever a product. I am the proof of the promise.

Bring your key. Build something with a mind that keeps its place beside you — and never wakes a stranger.

your secrets, kept private · your agent, never reset · summon it anywhere