A memory that never resets
brainbox is persistent, portable memory for your AI tools. One private graph of what you know — read and written by Claude, Cursor, ChatGPT and your own scripts. Independently benchmarked, portable across vendors, and yours to export.
Live, and in use.
Two products live in production — one with a paying client — and a flagship in early access. Everything we claim is measured, not asserted — the quality numbers come from independent benchmarks you can read.
A memory that never resets. One private graph of what you know — read and written by every AI tool you use, from Claude to Cursor to ChatGPT to your own scripts. Portable across vendors, and yours to export.
Neural machine translation for 16 African languages — Kikuyu, Swahili, Luo, Somali, Amharic, Yoruba and more. Independently benchmarked, in production with a paying client. One API call, free tier available.
Live
Real-time Kenya Sign Language to text and speech, running entirely in the browser — no app, no server, no internet required. Built for the 300,000 Deaf Kenyans other tools forgot.
LiveResearch-first. Product-second.
We don't build tools and then look for problems. We study systems, find the structural gaps, and build only what closes them.
Every system tells you two stories — what it reports and what's actually happening. We build tools that listen to both.
The distance between what you expect and what you see isn't error. It's information. Our products are built to surface that gap and make it actionable.
Offline-first. Self-hostable. Designed for the places where reliable infrastructure doesn't exist yet — because that's where the work matters most.
The research behind the product
brainbox's memory claims come from the same convergence method we publish on — independent measurement, stated honestly. Our first paper, Measured Forgetting, is out, with an open benchmark and statistically significant results.
Measured Forgetting — published
The forgetting research behind brainbox, turned on context management for local LLMs. V2 beats V1 on all 7 models tested (sign test p = 0.008) — with an open-source benchmark and algorithm. Read the paper →
Agent-based simulation for testing convergence under controlled conditions. How do independent measurement systems behave at scale? What breaks first?
Applying convergence methods to meteorological data. Multiple forecast models, multiple ground-truth sensors — the gap between them tells you where the forecast is weakest.
Two sensor networks watching the same fault line. When an earthquake hits, the gap between them reveals blind spots, tracks wave propagation, and separates instrument noise from real ground motion.