Memory, made legible

Enter the access code to continue.

That code is not right.

A deep-tech AI company, quietly building

AI the world can trust
starts with memory.

We are building the memory layer for the agent era: agents that know the people they work with, get things right, can prove why they acted, and can truly forget.

mission.asserted  ·  true since day one  ·  learned the hard way

The problem

Everyone is scaling capability.
Almost nobody is scaling trust.

Agents are being handed real work: money, contracts, care, decisions. And three failures follow them everywhere.

01

They forget.

Every session starts from zero. The person explains themselves again, and again, to something that claims to be a colleague.

02

They guess.

Old truths and new truths sit side by side in a context window, and the model picks one with confidence. That is where hallucinated "memory" comes from.

03

They can't explain.

When an agent acts on your behalf and someone asks why, the honest answer today is a shrug. That answer will not survive contact with a regulator, or a customer.

What we believe

A manifesto, with receipts.

Inside our kernel, every fact carries two clocks: when it became true, and when we came to believe it. We hold our own convictions to the same standard.

Memory is the difference between a tool and a colleague.true since the first conversation ever held  ·  learned watching users repeat themselves
An agent that cannot show why it believes something should not act on it.true since agents started touching money  ·  learned from every disputed decision
Forgetting is a feature. Provable forgetting is infrastructure.true since the right to erasure became law  ·  learned before it could not be retrofitted
Trust can be added later. Trust is the architecture, or it is nowhere.superseded  ·  the old belief is struck through, never deleted. that is the point.

What we're building

A kernel where trust is structural.

Not a vector store with better marketing. An event-sourced memory kernel where the properties everyone promises fall out of the architecture.

ledger

Nothing is rewritten

Every word lands in an append-only, hash-chained log. Ground truth stays ground truth.

append(event)
two clocks

Truth and belief, separately

What was true, and what the agent believed at the time. Corrections supersede; they never erase.

as_of("03-15")
receipts

Every answer shows its work

Each recall writes a receipt listing exactly what the model saw. "Why did you do that?" becomes a query.

receipt("RCP-7Q2F")
erasure

Forgetting, with proof

Redact the log, rebuild every projection, scan for residue. The chain still verifies. A proof, not a promise.

erase("subject") → verified
zero-token

Memory without burning the model

Remembering and recalling use no LLM at all. The one model call runs after the session, on a budget.

0 LLM tokens to remember

The demo, in miniature

Watch a memory form.

Four months with one person, compressed to half a minute. Every turn appended, every fact distilled with two clocks, every event chained. This is the real mechanism, in miniature.

conversation
memory · facts
ledger · hash-chained

Proof

It already runs.

This is not a whitepaper. A working kernel sits behind a live demo, measured on a public benchmark against the systems people actually deploy.

  • Zero tokens for memory. Remembering, indexing, and recall run without an LLM in the loop.
  • Far cheaper than replaying history into the prompt: published memory systems cut input tokens 55 to 75% against full replay and 40 to 55% against tuned RAG with caching. Run the savings calculator on your own call volume.
  • Milliseconds where pipelines take seconds, because retrieval is deterministic, not generative.
  • More accurate than LLM-pipeline baselines with the identical reader model, especially where facts change over time.
  • In the field with design partners running real agents in finance operations.
claims.compiled  ·  measured on a public bench  ·  receipts available in person
Chains of amber memory blocks pinned by dashed lines to the points on a grid where things happened
A tall lattice of light with an amber column of memory rising through it

Why now

The trust layer gets built once.

Agents are multiplying faster than anyone can supervise them. Regulation is arriving with penalties measured in percentages of global revenue, and it demands something no bolt-on can provide: lifetime records of what an AI knew and why it acted.

You cannot retrofit a memory. Whoever builds the layer that regulators, auditors, and users all trust will sit underneath everything built on top. That layer is being decided in the next few years, by a small number of people.

The invitation

Built by few, for the scale of everyone.

We are a small team with evidence discipline: every claim traced to a source, every design decision argued in writing, every benchmark number reproducible. We ship a real kernel, not slideware.

The scale we are building for is not a growth chart. It is the spec:

Billions of memories, tens of billions of causal connections between them, a million recalls a second, and a tail latency that cannot tell a deep chain from a wide fan-out. Auditable the whole way down, at the cost of ordinary cloud storage.ambition.asserted  ·  true since day one  ·  receipt to follow, measured at production scale
builders

Founding engineers

People who think in event logs and invariants, who want their systems work to define how a generation of AI behaves, and who find "it seems to work" an unacceptable sentence.

partners

Design partners

Teams running agents on work that matters, who want them personal, accurate, and defensible, and are willing to shape the platform with their hardest real cases.

If you're reading this, someone wanted you to. You already know how to reach us.