RTAI

Naturally fractal language models with persistent memory in their own weights.

RTAI is a from-scratch PyTorch research project exploring models that keep learning during operation. Its distinctive mechanism is self-modifying fast-weight memory, W <- gamma W + beta (v - Wk) k^T: constant-size state, no growing KV cache, and memory that can survive a process restart. The work is experimental, measured on modest hardware first, and reports negative results alongside successes.

Hardware and compute sponsorships

RTAI deliberately proves ideas on modest hardware before scaling them. Donations and time-limited access to neuromorphic hardware, AI accelerators, CUDA-capable GPUs, workstations, servers, ECC memory, or training-cluster credits can expand the experiments that can be tested honestly. Sponsorship does not buy favorable results, endorsements, or roadmap control; failed experiments remain part of the record. Organizations interested in supporting the research can open a GitHub issue titled Hardware sponsorship with the available resource class and access constraints.

Current evidence

Model or study	Size	Evidence	Status
SRWM symbolic PoC (rtai/)	0.82M	key-value recall persists across restart; ablation falls to zero	Validated in small
SRWM TinyStories LM (rtai/)	16.9M	storytelling plus persistent in-weight facts	Validated, limited
Fractal LM (fractal/)	108.7M active	weight-tied rule, timescale ladder, FineWeb-Edu pretraining	Partial run, not a finished model
Fractal recall study	about 4.6M	two-scale routing was strongest; empty-add neurogenesis failed	Study complete
Plasticity genome	small	one-pass sequence adaptation transferred scale, recall did not	Refuted for declared gate
Efficiency tournament	small	no candidate passed the promotion contract	Negative result
Growing Cortex	3.00M total / 0.57M active	16 compiled skills retain 76.2% across append and restart; 0% control hijack	Synthetic mechanism validated
Natural Cortex	104.55M stored / 65.90M active MoE	atomic English pipeline, deterministic dense/MoE gate, and local rank-8 teaching	MoE selected; resumable run paused at 25M tokens

The detailed measurements and falsification criteria are in docs/EXPERIMENTS.md. The non-negotiable design contract is VIBE.md.

Quick start

Python 3.12 and uv are required.

uv sync

Train the small symbolic model, then open its local memory visualization:

uv run python -m rtai.train
uv run python -m rtai.serve --ckpt ckpt.pt
# http://localhost:8000

Teach and query the same model from a terminal; its runtime memory is saved locally:

uv run python -m rtai.run --ckpt ckpt.pt chat

Run the FractalLM dashboard in truthful learn-from-scratch mode:

VIZ_LEARN=1 uv run python -m fractal.viz_serve
# http://localhost:8000

Run a tokenizer-compatible checkpoint with persistent chat and per-message memory ratings:

FRACTAL_CKPT=MODEL.pt VIZ_TOKENIZER=TOKENIZER.json VIZ_CHAT=1 VIZ_FEEDBACK=1 \
  uv run python -m fractal.viz_serve

Ratings 1..5 weaken, leave neutral, or consolidate a message into delayed-credit memory and a bounded W0 overlay. The mechanism, private runtime files, trainer queue, and falsification screen are documented in docs/EVENT_ALGEBRA.md.

VIZ_CHAT=1 alone provides persistent chat without enabling Event Algebra or W0 consolidation.

Append-only low-rank skill hemispheres, their meta-compiler, lifecycle, dashboard telemetry, and sequential falsification gates are documented in docs/GROWING_CORTEX.md.

The compiler-free English conversational preset, pinned data mix, deterministic dense/MoE gate, local teaching runtime, and preflight status are documented in NATURAL_CORTEX.md.

No checkpoint is currently published or stored in Git. Future model releases will use checksum-verified Safetensors on Hugging Face; see docs/MODEL_RELEASES.md.

Architecture

The repository contains two related lines:

• rtai/ is the compact SRWM proof of concept and language-model baseline.
• fractal/ applies one recurrent rule over depth and a ladder of memory timescales, with training, agent grammar, evaluation, persistence, and live architecture telemetry.
• docs/ is the experiment record, including failed approaches and their measured mechanisms.
• tests/ and fractal/tests/ cover kernel equivalence, streaming causality, persistence, safe checkpoint handling, and learning in small.

The dashboard is part of the architecture contract: it shows actual model geometry, data flow, and sampled signals from a real run. Telemetry is throttled and never adds a second training pass.

Development

uv sync --group dev
uv run pytest
uv run ruff check .
./scripts/audit-public.sh

Contribution requirements are in CONTRIBUTING.md. Security and runtime-memory privacy are documented in SECURITY.md.

Responsible status

RTAI is research software, not a production assistant or safety boundary. Small checkpoints can be wrong, brittle, and easy to destabilize. Persistent state may encode information supplied during operation and must be handled as sensitive local data. Capability claims are limited to the published test and experiment conditions.

License and governance

Source code is available under the Apache License 2.0, including commercial use and an explicit patent grant. Attribution notices are recorded in NOTICE. Project stewardship is maintainer-led as described in GOVERNANCE.md. Dataset and future model licenses are reviewed separately from the source-code license.