Evolution of the program

First contact. Small chat scripts, anomaly detection, toy agents — poking at AI the way you poke at a campfire.

• First agents & scripts archived

  ai_chat, anomaly-detection experiments, tiny "creature" agents (bunny, fluffy). No thesis yet — just learning the terrain.

Genetic algorithms and "chaotic breeding" of binaries. Mostly noise — but it taught the one idea that survived everything: structure can emerge from chaos if you let it loop.

• Chaos experiments (beast, boomXD, 1024doom) archived

  GA over binaries, self-referential loops, recursive feedback. Dead-ended as products, but seeded the recurring bet on emergence-from-recursion.

  ↳ evolved into: Fractal cognition (2025)

The obsession crystallized into real systems. Byte-level training from scratch, the fractal backbone, genetic architecture search, and inspectable local agents.

• byte_gpt → ForgeLM evolved Commodity Training

  From-scratch byte-level GPT training (124M–774M) on one 4090 matured into ForgeLM, the commodity-training stack.

  ForgeLM

• Fractal Neurons / LILA active Fractal Cognition

  A parameter-shared fractal backbone (~70M params, 65k+ runtime nodes), byte-level, MoE. The flagship intellectual core. LILA persona layer kept as speculative flavor.

  Fractal Neurons

• FNAS — fractal architecture search active Fractal Cognition

  Genetic search over fractal model genomes with LLM-as-judge scoring.

  FNAS

• QJSON Agents + YSON shipped Agent Runtime Systems

  Local-first agents whose identity and memory are plain, inspectable files; YSON persona format. FMM appeared here, independently, again.

  QJSON Agents

The recurring ideas got extracted, named, and — finally — benchmarked. Memory systems became the lab’s first proven thread.

• Centroid router benchmarked (Proof Drop #3) benchmarked Memory Systems

  Fired proof drop #2's oracle and bolted on the cheapest router that could work — one centroid per topic. When memory is separable it's nearly free (~0.002 ms/query) and recovers 98% of oracle recall at ~60× flat-scan speed; as topics overlap, realized recall tracks routing accuracy and collapses toward chance. The catch: proof drop #2 ran in exactly that hard regime, so its oracle win was real but uncashable by a trivial router. Routing quality gates the locality bet. Shipped route() in fmm v0.3.0.

  benchmark → · cite · fmm ↗

• FMM extracted, then benchmarked (Proof Drop #2) benchmarked Memory Systems

  The Fractal Memory idea — reinvented across four projects — pulled into a standalone library, then given topic-scoped retrieval and benchmarked: ~164× faster and ~3.7× more accurate than flat search at 128k items when the topic is known; recall 0.0 if misrouted. The locality bet, in semantic memory.

  benchmark → · cite · fmm ↗

• UFM extracted + benchmarked (Proof Drop #1) benchmarked Memory Systems

  Unified Fractal Memory generalized FMM from semantic to physical memory. First reproducible proof drop: runs a 24 GB model on a 23.5 GB GPU; full-GPU throughput when the working set fits; honest no-locality failure case.

  benchmark → · cite · ufm ↗

• NeuroArch active Cognitive Architectures

  The cognitive-architecture thread. A competent agent substrate today; NeuroArch is the background-cognition layer (idle reasoning, belief revision, crystallization) it does not yet fully run. Flagship long-term direction — next proof: make the cycles measurable.

  NeuroArch

• AEON: Living Worlds — open-sourced, embodied, documented active Cognitive Architectures

  The world-model project went public on GitHub and grew citizens that are embodied (positions, pathfinding, perception) and driven by a teacher-to-student liquid neural network. Shipped a technical whitepaper, and a devlog tracing a trainability bug in that network to a missing normalization — fixed with a LayerNorm so the student learns reliably across seeds. Framed as engineering; emergent behavior is not yet formally measured.

  AEON · devlog · aeon ↗

• agentos: from kernel to self-hosted agent platform shipped Agent Runtime Systems

  The terminal-native kernel grew into a mobile-first agent platform: a ReAct loop over 120+ models, ~26 real tools, human-in-the-loop permissions with a filesystem scope, runtime tool-forging in a hardened node:vm sandbox, and an ~88-skill catalog with a live, environment-aware system prompt. Engineering rather than a benchmarked result — the forging sandbox and skill catalog are adapted from the open-source @framers/agentos (Apache-2.0).

  agentos