ArborLM

Independent AI research

Teaching language models to grow a memory.

ArborLM asks whether a language model can store knowledge in a large, trainable memory it learns to navigate, and whether one small model can learn from several larger ones over time. It is a one-person research project, and everything it learns is published, including what fails.

What ArborLM is

Today's language models keep almost everything they know mixed into one block of weights. Updating that knowledge means retraining, and running a large model means moving huge amounts of data for every word. ArborLM explores a different shape: a model that keeps part of its knowledge in a separate memory, organised like the branches of a tree, and reads only the few pieces it needs.

Research

Open research

The research, at arborlm.ai. Each experiment is written down before it runs, with its pass and fail criteria, and the outcome is published as recorded. See it on GitHub.

Engineering

The model

The model being built from that research: a fast runtime, measurement and tools to look inside the memory while it learns. More below.

Goal

Learning from teachers

The long-term aim is a student model that can acquire, keep and combine skills from several open teacher models. That is a question to test, not a result.

The model: what is being built

The idea comes from the research: keep a pretrained model's weights frozen and give it an external memory that training writes into. The engineering challenge is to make that fast, measurable and safe to experiment with.

Runtime

Rust inference and training core

A runtime written for this architecture, checked number by number against a frozen reference before anything is built on it.

Memory

Trainable external memory

Attached to a small open pretrained model through a bypass, so the original model's behaviour stays available as a baseline.

Explorer

Seeing inside the memory

A visual explorer that shows which memory slots are read, how they connect and how that changes during training, with measured data kept separate from interpretation.

Telemetry

Honest measurement

Hardware and training telemetry with known overhead. Unavailable counters are reported as unavailable, and synthetic data is always labelled as synthetic.

Status: planning and early engineering. The source is not public yet, no model has been trained or released, and nothing here demonstrates the research hypothesis. A passing engineering test is never reported as a research result.

The plan

Four stages. Only the first is under way.

  1. Engineering foundation

    Runtime, telemetry, the memory explorer, a pretrained backbone running, and a trainable memory attached to it. In progress.

  2. Integrated prototype

    A student that genuinely learns new information into its memory, with reproducible training, checkpoints and recovery.

  3. Learning from teachers

    An experiment, not a promise: can one student acquire, keep and combine capabilities from open teacher models such as Qwen and GLM? A negative answer will be published too.

  4. Independent validation

    A separate milestone with independent evaluations against the teachers themselves. Finishing the stages above is not evidence that the approach works.

Where it stands

Honestly: early. Some ideas have shown promise in small, controlled studies; several have failed, and those failures are published next to the successes. None of the target capabilities is demonstrated yet, and nothing here is a product.

Why this matters for supporters: your support pays for careful experiments with honest outcomes, not for a promised result. The current state of every question is kept up to date in the ArborLM README.

Support the work

All experiments run on a single workstation, funded personally. Support goes straight into compute time, storage and the hours it takes to run studies properly.

Donate

Make a donation

One-off or recurring. Every amount helps keep experiments running.

Ask how to donate

Compute

Offer compute or hardware

GPU hours, cloud credits or a machine to borrow can make whole studies possible that do not fit on one computer.

Offer compute

Grants

Grants and sponsorship

Foundations, labs and companies interested in open research on model memory are welcome to get in touch.

Talk about it

ArborLM is not a registered charity or company. Support goes to an independent researcher, does not buy any rights or deliverables, and does not influence how results are reported.

Get in touch

Questions, support, compute offers or just curiosity: send a message and it reaches the researcher directly.

Or write directly to

hello@arborlm.com

You will get a reply from a person, usually within a few days.

Your message is emailed to the project owner and used only to reply to you.