For engineering teams
In private betaMemorie Cortex
The context management stack behind the agents you ship.
- Persistent memory with anticipatory retrieval
- Entity resolution and knowledge pipelines
- SDK for TypeScript and Python
Private beta
Build AI that remembers, learns and gets better over time.
Memorie gives your agents persistent, structured memory and active context management. Every interaction makes them more personal and more accurate, and you stop paying to re-send the same history on every call.
- Three stores
- vector, graph and file, queried in one pass
- Off the request path
- writes go to a queue, never blocking a turn
- Sealed by default
- contents encrypted with keys we do not hold
We are onboarding the first cohort by hand, a few teams at a time.
Illustration of the shape, not a measured run. We will publish real numbers with the harness and the method once the beta closes.
What is Memorie?
Memorie is the memory layer that keeps your data encrypted, including from us, so the agents you build and the assistants you use can finally hold on to something. Two products share one engine. One is open to beta teams today, the other is not built yet.
For everyday AI work
Coming later this yearMemorie Keep
A private context layer that follows you between tools.
- Encrypted personal vault, closed to us by default
- Holds preferences, work history and open threads
- Carries context across every assistant you use
01/The problem
Why memory, and why the usual answers fall short.
A bigger context window moves the cliff, it does not remove it. Vector search returns things that sound alike rather than things that matter. Flat files stop scaling the day two agents write to them. And the platform vendors are not going to own your memory on your behalf.
Recency bias
The middle of a long thread quietly stops being read.
Cost curve
Re-sending history bills you again for what you already said.
No structure
Raw transcripts cannot be queried, corrected, or expired.
02/How it works
How Cortex actually works
A turn does not drop into a database. It gets ingested, its meaning is pulled out into structure rather than stored as raw text, and it is written across a vector, graph and file store, asynchronously, so writing never blocks your agent. Retrieval then nets across all three at once.
- 01
Ingest
async write
A turn lands on the queue, not in your request path.
- 02
Extract
structure, not raw text
Entities, claims and relations, resolved against what is already known.
- 03
Store
vector · graph · file
Three stores, each holding the shape it is actually good at.
- 04
Retrieve
anticipatory
One net across all three, fetched before the agent asks.
- 05
Consolidate
background cycles
Merge duplicates, promote what recurs, expire what went stale.
Vector
semantic similarity
Graph
entity relationships
File
documents and raw source
03/Roadmap
What exists, and what does not yet
We would rather you knew which parts are real before you write any code against them. This list is the honest version, and it changes as things ship.
In the beta today
- Cortex SDK for TypeScript and Python
- Async ingest with the full extract and store pipeline
- Retrieval across all three stores in one call
- Per-namespace scoping and hard delete
Next
- Published benchmark run on an open harness, with the method
- Consolidation policies you can configure per namespace
- Self-serve signup and usage dashboard
- SOC 2 Type II audit begins
Later
- Memorie Keep, the personal vault
- Bring your own cloud deployment
- SSO and audit export
We have not published benchmark numbers yet. When we do, the harness, the dataset and the run will be public alongside them, because a memory accuracy claim with no method behind it is not worth reading.
04/Security
Encrypted, including from us
Client-side keys
Vault contents are sealed before they leave your process. We hold ciphertext.
Scoped retention
Per-namespace expiry and hard delete, enforced in the consolidation cycle.
Isolation by tenant
Separate stores per workspace, no shared index, no cross-tenant recall.
Audit trail
Every write, read and forget is attributable to a caller and a reason.
We are not certified yet. The SOC 2 Type II audit starts once the beta closes, and we will say so here rather than putting a badge up early.
Private beta
The difference between a demo and a product is memory.
Tell us what your agent forgets and we will tell you honestly whether we can help yet. Free for the whole beta, and we onboard a few teams at a time.