Data sovereignty
A memory pool is the most concentrated form of your organization's IP. Code shows what you built; memory records why — the decisions, reversals, constraints, and context that never make it into the repo. That concentration is exactly why a memory layer must be able to run at whatever level of sovereignty your risk profile demands. Anamnesis is built as a ladder: start where convenience matters, graduate as sensitivity grows, without changing products.
The tiers
| Tier | Where content lives | Who can read it | Status |
|---|---|---|---|
| Hosted | Managed service; per-user encryption at rest, no master key | You; operators with production access (stated plainly — see Trust Center) | Live (private beta) |
| Client-held keys | Managed service stores ciphertext it cannot read; keys never leave your machines | You alone. Key loss means data loss — no escrow, by design | Published phased roadmap |
| On-premises | Your hardware, entirely. Memory, consolidation, and inference all run locally | You alone; nothing leaves the building | Design partner conversations |
Each tier is honest on its own: we state what the current tier does and does not protect against, rather than marketing the top of the ladder while shipping the bottom.
The one-way mirror
The piece that makes shared memory compatible with IP containment. A local or guest client can be granted read-only access to a shared pool:
- Context flows out — the local model draws on the team's collective memory on every turn.
- Nothing flows back — the guest's writes stay on its own machine. Its generations can never reach the shared pool, and therefore can never be pulled into a third-party model's context.
The guarantee lives in the credential, not in client discipline: read-only scope is enforced at the API layer on every route (architecture).
Local inference
On the sovereignty ladder, where the model runs matters as much as where the data sits. The on-premises tier routes inference per session: a local model by default — sensitive queries answered without anything leaving the machine — with hosted models available only when a session explicitly opts in. Consolidation (the pipeline that turns sessions into durable memory) runs locally in this tier, which also removes the hosted tier's LLM sub-processor dependency.
Who this is for
- Regulated industries — where "our vendor can technically decrypt it" is a failed audit answer. The ladder gives a compliant landing spot without giving up the product.
- IP-sensitive engineering organizations — where the concern isn't compliance but competitors: decision memory is the last thing you want leaving the building.
- Teams that want leverage without lock-in — memory that is model-agnostic and exportable is memory you own. Switching assistants shouldn't mean amnesia.
The straight answer on where we are
Hosted is live and honest about its trust boundary. Client-held keys is a published, phased roadmap — each phase shippable and true on its own. On-premises runs today as a working reference implementation and is being shaped with design partners; if your organization needs this tier, we want to talk: hello@jiashley.com.