Research

Open research,
quietly kept.

Three threads of work the studio is doing in public. The physics, the CS history, and the memory.

I. The bipartite reading of Einstein–Maxwell

● Under peer review · Foundations of Physics (Springer Nature) Submitted 2026-05-10 SNAPP 6ee9ff0b-fc04-…

The studio's primary physics output to date. A reading of the Einstein–Maxwell system that organises light–gravity coupling into three orthogonal channels — gravitoelectric, gravitomagnetic, and Λ-cosmological — rather than collapsing them into a single tensor expression. Each channel is examined in its own right; together they form a complete three-channel taxonomy.

A 54-round computational audit (v1 → v2.57) has held the conjecture frozen at 0.5 / 11 SPECULATION promotion conditions met — the correct answer when the open problems are open. — from the corpus, Apr 2026

The interactive 3D companion lets a reader walk through each channel visually. Built in Three.js, lives at /theory.html. The paper itself is awaiting editor decision by ~June 10 2026.

II. The Algorithm Lab as research

● Live 39 algorithms · 2,300 years 3D · synced pseudocode · live STATE

It's a product. But it's also a piece of history-of-CS pedagogy: every algorithm carries its year, its author, and a brief origin story. The chronological view starts with Euclid's GCD (300 BCE) and ends with Tim Sort (2002). What started as a homepage demo became, quietly, the most comprehensive visual archive of mainline CS algorithms on the open web.

The studio's research thesis on it: interactive figures beat passive video at every measure of learning retention worth measuring. Bret Victor's *explorable explanations* applied to a curriculum domain where the standard pedagogy stopped evolving in the 1980s.

III. aimem as a memory-architecture implementation

● Open source · MCP-ready github.com/LeKCei/ai-memory Used in production

A working markdown implementation of seven separate research-paper architectures. Not a thought-piece. Not a framework. A single canonical store of personal context, with per-tool adapters built from it, that you can clone and run today.

The interesting part is what the existence of aimem claims: that the right memory layer for an AI assistant is not a vector database with a magic-retrieval API. It is plain markdown files, organised by a research-grounded taxonomy, with a build step that emits per-tool views. The complexity goes into the structure; the storage stays inspectable.

Source papers
Sumers et al. — Cognitive Architectures for Language Agents (CoALA) · Princeton, 2023
Packer et al. — MemGPT: Towards LLMs as Operating Systems · UC Berkeley, 2023
Xu et al. — A-MEM: Atomic Memory for Language Agents · NeurIPS, 2025
Fang et al. — MEMP: Procedural Memory Distillation · Aug 2025
Zhong et al. — MemoryBank: Enhancing LLMs with Long-Term Memory · AAAI, 2024
Park et al. — Generative Agents: Interactive Simulacra of Human Behaviour · Stanford, 2023
Mem0 framework + TiMem reflection schedule · 2024–2026

What's next.

Three more threads quietly underway, in roughly the order they'll publish:

All of it will land on this page as it ships. RSS feed for the impatient; Changelog for the curious.