Your AI knows everything about the world. Nothing about you or your business.
Geometrical is the memory runtime for AI. It captures what happens as it happens, compiles a private graph inside your perimeter, and lets every model see only what it needs.
Models are rented.
Memory is owned.
Public graphs made assistants useful. Private graphs are next.
Everyone will need a private graph.
People, families, teams, companies, and agents.
What matters never reaches a system.
Decisions and their reasons live in people. When they leave, it leaves.
Every prompt overexposes.
Models see far more of you than the task needs. That gap can be computed.
Agents will share memory.
Without permissions between them, a private graph stops being private.
Memory compounds — or rots.
The accumulated graph is the asset. Keeping it true is the work.
Context is compiled inside your perimeter. Only the minimum leaves.
Inputs become four planes of memory. Threads run through them to one lens: the minimum context the task needs, and the minimum exposure it may carry.
Hover any part of the diagram.
Memory is written when it happens. Not reconstructed later.
A decision captured in the moment keeps its who, when and why. Rebuilt from documents a week later, it keeps only the what.
Read the architecture →
Every AI memory has to answer four questions. Most can't.
What kind of graph?
Most AI memoryDocuments and entities. It remembers who.
GeometricalEpisodic, causal, temporal and relational, with roles and vertical ontologies. It remembers why.
Who keeps it true?
Most AI memoryNobody. It grows, then it rots.
GeometricalWe do. Dedup, sanitise, version and verify — without seeing the content.
What leaves your perimeter?
Most AI memoryWhatever was retrieved, in clear text.
GeometricalThe minimum viable context, with sensitive data transformed locally before inference.
How is it measured?
Most AI memoryRecall on a benchmark.
GeometricalExposure Ratio, MVE Score, token reduction, p50/p95 latency, graph health.
And four more nobody asks yet.
Nexus: deny-by-default, purpose-bound, revocable, logged.
.aix, an open and portable format.
Less: compiled context means fewer tokens per answer.
Yes. The model is replaceable; the graph is not.
The question isn't graph or no graph. It's which of three levels you're on.
Ask your company“Why did we drop that supplier?”
The decision, who made it, what triggered it, what happened next.
| No memory | Flat retrieval (RAG) | Governed episodic graph | |
|---|---|---|---|
| Someone leaves | What they knew leaves too. | Their files stay. Their context doesn't. | Their decisions and relationships stay. |
| Sensitive data | Pasted into prompts by hand. | Whole chunks sent to the model. | Transformed locally. Only the minimum leaves. |
| Cost | Long prompts, repeated. | Redundant chunks. | Compiled, deduplicated context. |
| Audit | None. | Hard to know what the model saw. | Who saw what, for which purpose. |
| Many agents | Each starts from zero. | Shared index, no permissions. | Permissions per purpose, revocable. |
| Change model | Start over. | Re-index. | Nothing changes. |
| Over time | Zero. | More volume, same understanding. | Compounding: each episode strengthens the graph. |
What a model needs and what it's allowed to see are not the same thing.
| Minimum Viable Context (MVC) | Minimum Viable Exposure (MVE) | |
|---|---|---|
| The question | What does the agent need to do the job well? | Which private data does the model need to see? |
| Optimises | Usefulness and cost | Privacy and compliance |
| Measured in | Tokens | Tokens weighted by sensitivity class |
100%
Everything retrieval found goes to the model.
A short context can still overexpose.In the paper's worked example, removing unnecessary context cuts tokens by 92.9%. Weighted exposure is still 203× the minimum.
Geometrical is MVE-first.It compiles the minimum viable context, then enforces the exposure limit before anything reaches the model.
No new app. No new habit.
- Memory shows up inside the tools you already use.
- Capture is passive and continuous. Set the perimeter once, then forget it's there.
- Each surface comes pre-configured with legally permitted sources only. At work: corporate systems approved by IT. Never screens or keystrokes.
- Supplier dropped after three late deliveries in a row.
- Decided by the ops lead, approved by finance.
- Replacement onboarded the following month.
3 facts · sensitive fields kept local · purpose: reply to finance
One memory runtime. For a person, a family, a team, a company, or an agent.
SelfFor you and the people close to you.
Measured, not promised. Every number here has a date and a machine.
Causal edges
Self v1.1 · founder's devices · zero manual input
Entities
Same graph
Roles inferred
Same graph
Ingestion sources
Same graph
Latency, p50 / p95
2015 Intel MacBook
What we haven't proven yet.
- Fully local inference is not fast enough yet on older hardware.
- Our empirical MVE benchmark (v1.1) is in progress, pre-registered.
- Our first enterprise pilot is running.