Our thesis
AI that builds on its own reasoning.
Today’s agents start every session from zero. We think the next stage is AI that remembers what was decided and why, tests new claims against evidence, and builds on what survived. This page explains why we believe that, and why it takes two products.
Where this goes
The stage after agentic workflows is compounding.
Each stage arrived when something new made it possible. The next one needs memory that keeps the reasoning, not the transcript.
A model that talks
Fluent, and starting fresh every time.
Bigger models and windows
More in one pass, and more to re-read.
Agentic workflows
Where the industry is, and where it meets a ceiling.
Compounding AI
Work that builds on what was already reasoned out, because the reasoning is kept, not the transcript.
Human-level reasoning
Memory and analysis working as one.
Why there is a ceiling
Every session starts from zero. Compaction keeps the tail of a transcript and throws away the reasoning. Cost grows with what gets re-read rather than with what was learned. Teammates re-explain the same settled decisions to the same agent.
What reasoning memory keeps
Why something was decided, what was decided, and how at a high level. Not transcripts, and not embeddings of everything. It is hierarchical rather than a flat graph, running from mission to decision to detail, because that is the grain a person moves through. A graph has no grain.
What that buys
A working set of a few thousand tokens instead of re-reading hundreds of thousands. Reasoning that merges with what came before instead of replacing it. It also runs 2–3× cheaper at equal recall once a session is long enough for a full transcript to weigh.
Why one company
Reasoning has two faculties. We build one product for each.
Remembering holds what you know and why, at the right grain, and retires it when it turns out wrong. Analysing derives new conclusions from evidence: form a hypothesis, test it, keep what survives.
Neither compounds alone. Memory without analysis is recall, and can only return what it was given. Analysis without memory starts over every time, re-deriving the definitions, re-finding the last result, re-arguing the settled decision. Wired together, analysis leaves assertions behind and memory hands them to the next analysis. That is what compounding means, mechanically.
Remembering
Assertion Memory keeps assertions with their reasoning, and supersedes them when reality moves. It works inside the coding agent you already use, or as our own agent.
Analysing
Assertion Analytics makes and tests assertions against your data, then records what held alongside the outcome that followed.
Our mission
Human-level reasoning in AI, one assertion at a time.
Built from two faculties, on one foundation, so that what a company learns keeps working for it.