FMM (Fractal Memory Matrix / Memory-Mapped) → UFM (Unified Fractal Memory). The FMM idea independently reappeared in Fractal Neurons, QJSON Agents, fnn_test, and NeuroArch before being extracted; UFM generalized it from semantic memory to physical memory (VRAM/RAM as one pool).
yan@yandesbiens:~$ cat research_program.md
The Research Program
Independent AI research by Yan Desbiens, conducted at Éthiqueia Québec inc. — one evolving program, not scattered repos.
Capable AI you can own — on commodity hardware — through self-similar architecture and memory that organizes itself.
maturity: conceptprototypeshippedbenchmarkedpublished · a thread is only “benchmarked” once a one-command repro exists.
Memory Systems
benchmarkedSelf-organizing memory and memory paging — memory that grows, prunes, and pages itself.
lineage
lead artifact
evidence status
Three reproducible proof drops (2026-06-27). #1 (UFM): runs a 24 GB expert bank on a 23.5 GB RTX 4090 where baseline OOMs; within ~1% of baseline throughput when the working set fits; ~240× faster than naive offload. #2 (FMM): topic-scoped retrieval ~164× faster + ~3.7× more accurate than flat at 128k when the topic is known; recall 0.0 if misrouted. #3 (FMM router): a near-free centroid router (~0.002 ms/query) recovers 98% of oracle recall at ~60× flat-scan speed when topics are separable, but realized recall tracks routing accuracy and collapses toward chance as topics overlap. All three converge on one finding: locality pays only when you can address the right region.
next proof needed
A learned topic router that pushes the routing-accuracy crossover into harder (overlapping) regimes where the cheap centroid router fails — measured end-to-end against the centroid floor (proof drop #3). Plus: real-embedding retrieval corpora; training-time paging (autograd-safe) + OffloadedAdam memory curves.
Cognitive Architectures
prototypeSystems that keep reasoning between prompts — internal state, world models, idle cognition.
lineage
NeuroArch — a cognitive architecture with idle cognition, self-revising beliefs, and memory crystallization, running on a competent local-agent substrate · AEON (an AI-governed deterministic world).
projects
- NeuroArch The agent substrate is already competent today. NeuroArch is the architecture on top — idle cognition, belief recalibration, dream-phase crystallization — which it does not yet fully run. The thread is about giving a working agent those background cycles and then measuring them.
- AEON · repo ↗ AI-governed deterministic world, now open-source and built in public, with embodied citizens driven by a teacher-to-student liquid neural network. Documented in a technical whitepaper and an ongoing devlog.
lead artifact
evidence status
Running systems, documented in public. AEON is open-source with a technical whitepaper and a devlog (e.g. tracing a liquid-network trainability bug to a missing normalization and fixing it). Internal/emergent behaviors are not yet formally measured, so all capability claims stay marked speculative — this thread is framed as engineering, not results.
next proof needed
Measurable, reproducible behaviors: memory growth over time, belief-revision traces, idle-cycle throughput, and long-run determinism/drift soak tests — before any public capability claim.
Agent Runtime Systems
shippedLocal-first agents whose identity and memory are plain, inspectable files.
lineage
QJSON Agents + YSON (persona/state format) → Alicia (local autonomous agent) → agentos (live kernel). (OpenPaw exists as background tooling, not a core research pillar.)
projects
lead artifact
evidence status
Code others can run; agents whose identity + memory are plain files. The YSON spec and an inspectable-memory demo are not yet written up.
next proof needed
A published YSON spec; a reproducible demo of inspectable, file-based agent memory compared against opaque memory stacks.
Commodity Training
benchmarkedTraining and runtime experiments on a single consumer GPU — raw bytes to trained model.
lineage
byte_gpt → ForgeLM → the 4090 capacity planner / autopilot. UFM (Memory Systems) is the bridge that lets this thread scale past 24 GB.
projects
lead artifact
evidence status
Measured training envelope (VRAM, throughput, tokens/day) for the full 30M-500M family on one RTX 4090, re-verified 2026-06-29, plus a real 120M loss curve (9.65 → 2.78 over 9.04B tokens). One-command reproduce.
next proof needed
Model-quality numbers from a full run — held-out perplexity and HellaSwag/ARC via `forge eval` — to put a quality axis next to the cost axis.
Fractal Cognition
prototypeSelf-similar architecture and reasoning — parameter-shared fractal backbones.
lineage
Fractal Neurons (the flagship) → fnn_test (fractal neurons as conversational agents) → FNAS / FNAS-V2 (genetic search over fractal genomes). LILA is the persona/emergence layer — treated as inspiration, explicitly speculative.
projects
lead artifact
evidence status
Rich and documented, but NOT yet benchmarked. All architectural-advantage claims are marked speculative until measured.
next proof needed
A controlled, small-scale comparison of the fractal backbone vs. a parameter-matched transformer (loss/perplexity at equal params + equal compute).
// the evidence bar
Nothing ships with a performance claim unless it is (a) reproducible from a one-command script in-repo, or (b) explicitly tagged speculative. Visionary concepts are inspiration; engineering is proof. We earn credibility on the verifiable threads to buy the right to publish the speculative ones.
→ See it all connected in the knowledge graph. Every reproducible output is a citable asset: see the citation registry. Follow new work by proof-drop RSS or the newsletter.
yan@yandesbiens:~$ subscribe --proof-drops
Follow the program.
Each thread advances by proof drop. Get them as they ship — benchmarks, architectures, field notes.