OnMemory
Request access

Engineering velocity at scale

Several models. One codebase. One record.

OnMemory lets a fleet of AI agents, on whichever models you choose, build, review and sign off on the same codebase from one shared, verbatim memory. Remove the bottlenecks. Keep the quality controls.

In closed test with invited developers · Public release December 12, 2026

buildsreviewssigns offreadsClaudeVENDOR ACodexVENDOR BGrokVENDOR COpen sourceANY MCP CLIENTOne verbatim recordattributed · deterministic

Every model writes to and reads from the same record. No vendor owns it.

Why now

One instruction. Hundreds of actions. Several models.

Software development has changed shape. A developer no longer writes code line by line; a developer directs a fleet. The bottleneck has moved from writing to reviewing, and the quality controls have not kept up.

The pace

A single instruction to a fleet of agents now sets off hundreds of actions across several models, and that number only grows.

The blind spot

No human reviews code at that pace, and a model checking its own output carries its own blind spots into the review.

The answer

The only review that scales is an independent model, from a second vendor, reading the same record and seeing what the first one missed.

How it works

Multi-model execution on a shared record

The same loop your best engineering team already runs, at the speed of a fleet: one party builds, another reviews, decisions are recorded, and nothing is lost between models.

Step 1

Specify

You write the specification. OnMemory keeps it verbatim, as the reference every model works from.

You

Step 2

Execute

Agents on the models you choose implement it, as many at once as the work needs, all writing to one memory.

Any model

Step 3

Review

A model from a different vendor reads the same record and reviews work it did not write, with none of the author’s blind spots.

A second vendor

Step 4

Sign off

Decisions, corrections and approvals are recorded and attributed, so the history of the codebase is provable, not remembered.

On the record

Why it only works on a shared record

Four properties no provider’s memory has

Every model vendor ships a memory of its own, locked inside its platform. Multi-model execution needs the opposite: one record that every model can trust, read equally, and never own.

V

Verbatim

The original record is the record. Nothing is summarized, compressed, altered or quietly discarded, so what a reviewer reads is what the builder did.

A

Attributed

Every record names the model that wrote it, is content-addressed, and is chained into a hash-linked audit log.

D

Deterministic

The same question returns the same answer, for every model, on every machine. Retrieval runs on the E8 lattice in fixed-point arithmetic, so there is no drift.

O

Owned by no vendor

Served to every model over MCP. It belongs to no harness, no control plane and no model provider. It belongs to you.

Memory is the layer. Multi-model execution is the point.

OnMemory is not an inference product and does not compete with the model providers. It is the record their models share, and the reason you can switch between them.

Proof

We run it on ourselves

Since December 2025, our own fleet of AI agents has built OnMemory using OnMemory. Every figure here is traced to that record.

22,800

commits since December 2025

2,000,000

memory records written by the fleet

3

vendors’ models working on one memory

30+

agents at once, with over a hundred subagents at peak

63

Defects found in our own platform in six days

One specification, a five-stage review in which every artifact is checked by a party that did not author it. Every blocking defect was found by a reviewer that did not write the code; not one by an author reviewing its own work.

That register is why independent, multi-model review is the only quality control that scales. The summary is public; the full register is in our data room.

Works with

Every model you already use

Built on the Model Context Protocol, the open standard for connecting AI tools. Add OnMemory once and every client shares the same record.

ClaudeClaude CodeOpenAI CodexGrokOpenCode · open-source modelsAny MCP client

Pricing

Start free. Pay when your fleet does.

Planned pricing for the public release. Final pricing is set with the first customers.

Every tier keeps the record verbatim and attributed, and every tier exports in full. The per-developer fee is a fraction of what the same developer’s company already pays for the models.

Be first in line

Put your fleet on one record

OnMemory is in closed test with invited developers today. Tell us about your team and how you build, and we will reply when your account is approved.

Public release

December 12, 2026

Free tier at release. Team pricing at release. Multi-model review in 2027.