Barcelona · Est. 2026
The inference economy runs on context. Human syntax is not the native language of machines. Geometrical is the memory substrate on which agents operate.
01 / Thesis
Agents need to know who they are working for — their goals, decisions, criteria, roles, and history — not just the last message in a thread. Today, that context is improvised, duplicated, and forgotten on every new conversation.
Geometrical builds the primitive that was missing: a persistent, structured, privacy-native memory substrate. Local by architecture. Machine-readable by design.
All surfaces share the same engine — they differ only in subject and governance model.
S-01
Personal Memory Substrate
The episodic memory of the individual. Captures intention, decisions, and criteria — not just facts. Operates across roles: founder, parent, partner, learner. Zero interaction required.
S-02
Institutional Memory Substrate
The same 9-dimensional model applied to collective entities. Organizational knowledge that survives team changes, restructuring, and time. Enterprise deployment.
S-03
Permission Protocol
Governs what memory dimensions are shared between a Self and a Soma, and between Self instances. Federated graph with user-controlled perimeter definitions. Open source.
S-04
Operational Memory Substrate
Memory substrate for autonomous agents. Agents running within the Nexus permission boundary can read the relevant dimensions of a user's identity graph. Humans control what context is made available.
02 / Core Engine
Five subsystems, shared across all surfaces. The architecture is what makes Geometrical infrastructure rather than a product.
ENGINE · 01
Knowledge Compiler Dedup LSH + JSON Distiller
Converts raw episodic signals into structured, deduplicated memory units. Locality-sensitive hashing eliminates duplicates before storage. JSON distillation produces machine-native context.
ActiveENGINE · 02
PII Protect Layer Privacy by Architecture
Sensitive entities are transformed before any LLM interaction. Privacy is a mathematical property of the data, not a policy applied after the fact. Patent filed.
PatentedENGINE · 03
Semantic Version Control Git for memory
Every memory write is a versioned, staged commit. Beliefs are not overwritten — they are evolved, with full audit trail. Prevents belief corruption in long-term identity graphs.
ActiveENGINE · 04
Permission Engine Nexus governance layer
Controls what memory dimensions are readable by which surface or agent. Permissions are explicit, user-defined, and revocable. The boundary between what an agent knows and what it may know.
In DesignENGINE · 05
Predictive Context Engine Role-routed retrieval
Infers the active role from behavioral signals and routes memory retrieval accordingly. No explicit declaration needed. Founder context and parent context do not bleed into each other.
Active03 / Self
Personal Episodic Memory Substrate
"You don't remember how you breathe.
You won't remember how agents know what they know.
That is Self."
Self is not an app. Not an assistant. Not a hub. It is the memory layer that makes every agent you interact with coherent over time — across roles, across devices, across contexts.
It captures what matters: intention, criteria, decisions, patterns. Not transcripts. It operates passively across eleven ingestion sources. It requires zero ongoing interaction.
04 / Publications
05 / Contact
We are building with a small number of partners. If you are working on the agent stack, investing in infrastructure, or want to pilot Soma inside your organization — reach out.
hello@thegeometrical.com