Built by someone who has shipped under an auditor.
flmnt exists because agents act on decisions that have already been reversed, and nothing in the memory category was measuring it. The company is one founder, twenty years of delivery at scale, and a decade of shipping software that had to satisfy a regulator.
Mike Dyer.
Twenty years shipping at scale, ten of them leading regulated delivery for enterprise and Fortune customers. The SOC 2 observation window flmnt is inside right now is not a first attempt at compliance — it has been the day job for a decade.
Software engineering leader at AstraZeneca — FDA, GxP and EU-governed patient monitoring and clinical-trial release software, with teams across seven countries on multi-cloud infrastructure.
Now VP Technology at Wits Innovation Lab, running delivery where FDA, HITRUST and SOC 2 are designed in from day zero. Prior programs under HIPAA, PCI/DSS and FedRAMP.
Twenty years building for Toyota, CBS, Warner Brothers, Lehman Brothers, VW and Target. BCG Digital Ventures alumni. Award-winning work — FWA, Google Sandbox, Ad Age.
CTO at Spylight through acquisition. First lead engineer at Tradesy, scaling 10 → 50 through Series A. Six years as a principal taking early-stage companies to product-market fit and through raises.
The error nobody was measuring.
Ten years of regulated delivery teaches one habit above all others: assume the record will be read back by someone who was not there. That is the habit the memory category skipped.
An agent would act on a decision the team had already reversed. The reversal was written down; it just looked exactly as current as the thing that replaced it. On a regulated program that is not an inconvenience — it is a finding.
Every tool in the category inferred its structure from prose after the fact. None of them published a result for a superseded decision, because an inferred graph cannot reliably express one. So we built the test, measured it, and published the number.
flmnt is the conclusion of that: memory where the agent authors typed causal events as it works, so the relationship between decisions is recorded rather than guessed. It runs in production today, and every claim on this site traces to a benchmark anyone can re-run.
A product of the Complai ecosystem.
flmnt grew out of Complai, a platform built for regulated and enterprise customers — the environments where a decision has to be provable, not just remembered. It is now offered on its own.
Every architectural decision behind the platform — including every reversal — lives in flmnt and is retrieved from it daily. Over 20,000 memories across the application. It is our first and most demanding customer.
Three commitments.
Small company, published numbers, no roadmap theatre.
Measured, or not claimed
Every number we publish comes from a benchmark that ships with the release. If we cannot measure it, we do not say it.
Authored, not inferred
No extraction model sits between what happened and what gets stored. The agent declares the relationship as a typed argument.
Built for the audit
Append-only and replayable from the first commit, because the customers who need this most are the ones who have to prove it.
Tell me what your agents keep getting wrong.
Founder-led, so the reply comes from me. Partnership, diligence, or a question about the benchmark — all the same inbox.