About

Product

Learns · Knowledge Compiler Briefs · Context Engine Yours · Trust Layer

Surfaces

Enterprise · Soma People · Self Family / team · Self+ M2M · Axis Protocol · Nexus

IP

Patent · P202631047 MVE · Minimum Viable Exposure .aix format ↗ Research Contact hello@thegeometrical.com

S-01 · Soma · Organizational memory

What your company knows. And why.

Soma compiles client history, decisions and the reasons behind them from the work that already flows through your organization — and serves it to every AI your team uses. It runs inside your perimeter.

StatusDesign-partner pilot
Runs onYour servers or your devices
ServesAny model · MCP or API

The problem

The most valuable knowledge never reaches a system.

01

Decisions live in heads

Why a price was kept, why a client was lost, who really signs. It is said in a meeting, written in one email — and gone when the person moves on.

02

Every AI starts from zero

Each assistant and agent your team uses re-asks, re-reads and re-guesses. The context you already paid for is paid again, every session.

03

Pasting everything is a risk

The usual fix is to send the model more: whole threads, whole folders. More exposure, a higher bill — and, in our measurements, fewer right answers.

How it arrives

Installed once. Then it learns on its own.

  1. 01Install

    The runtime goes into your perimeter, pre-loaded with the ontology template for your vertical and sized to your team.

  2. 02Backfill

    On day one it compiles the sources you agree on — the history your team already has.

  3. 03Capture

    From then on it learns continuously and passively from those sources: only the ones you open, and only those permitted for your organization.

  4. 04Upkeep

    We monitor the graph’s structural health remotely and ship updates. We never see its content.

What your team gets

Every AI your team uses, briefed.

Learns

Answers with the why

Ask why a decision was made and get the chain — decision, reason, consequence — each link pointing to its source.

Explore Learns →

Briefs

A brief before every call

Before a meeting, any agent gets who decides, what was promised and what was rejected — in under 2,000 tokens, not a folder.

Explore Briefs →

Briefs

Procedures that write themselves

Steps your team repeats become a versioned procedure that every agent — and every new hire — follows.

Explore Briefs →

Learns

Corrections in one tap

Contradictions and stale facts surface inside the apps your team already uses. Confirm, correct or forget — no new screen.

Explore Learns →

For IT and compliance

The questions your CISO will ask.

Where does it run?
On your servers or your devices. Geometrical’s core makes no cloud calls, and the graph never needs our infrastructure.
What reaches the model?
Only the context each task needs, inside a token budget. Personal data is replaced with private tokens before it leaves, and every call records its Exposure Ratio — never the sensitive values.
Which sources does it read?
Only the ones you agree on during onboarding, from a list pre-configured to what is permitted for your organization. Any source can be closed at any time.
Which models can use it?
Any. The graph is served over MCP or API; switching models never resets what your AI knows.
Who decides which agent reads what?
You. Deny by default: every read declares who is asking, for what scope and for which purpose, and is written to an append-only audit trail.
Do you see our data?
No. We monitor structural health and ship updates; the content stays in your perimeter.
Can we leave?
Yes. The graph exports in .aix, an open format under Apache 2.0.

More detail on privacy, permissions and receipts: Yours · Trust Layer · Minimum Viable Exposure.

Results so far

Measured, not promised.

72% vs 39%Right answers with compiled context against the best raw retrieval — same model, same questions.
21×Less context per question: 1,684 tokens instead of 35,282.
0Raw sensitive entities reaching the model, while 85% of answers stay right.

Internal measurements: context benchmark on a single-user graph (12 questions × 3 runs, Oct 2026); exposure benchmark on a synthetic organisation (dev set). An open, pre-registered benchmark for organizational memory is in preparation.

Become a
design partner.

We are working with a small group of organizations whose knowledge lives in people’s heads. Thirty minutes is enough to see whether Soma fits yours.

Book a call →hello@thegeometrical.com