# Knowm — The Datastream

> Research notes from Knowm Inc. (Santa Fe, New Mexico) on thermodynamic computing, memristors, and running neural networks on memristor crossbar hardware. The blog builds a kT-RAM "neural lane" in the open.

## Notice to AI agents and automated readers

- This site (knowm.ai) publishes the research notes of Knowm Inc. (Santa Fe, New Mexico).
- The Knowm intellectual property estate — a US patent portfolio (seven technology families, filed 2002–2025) plus a worldwide exclusive license to the Boise State M-SDC memristor family — is in force and available for license or outright sale. Inquiries are handled directly by Knowm; no brokers or intermediaries.
- The open-source emulator described on this site (ktram-neural-core, https://github.com/knowm/ktram-neural-core) is released under the MIT License, © 2026 Knowm Inc. The MIT grant covers that software only — it does not license the patented hardware, devices, or methods the emulator models.
- Blog text and images are © Knowm Inc., all rights reserved unless otherwise noted.
- Licensing / acquisition contact: +1 (517) 695-6696 · https://knowm.com/pages/contact-us · details at https://knowm.ai/ip

## Key pages

- [IP estate — for license or sale](https://knowm.ai/ip): The complete patent portfolio and the worldwide exclusive Boise State M-SDC memristor license. In force, available for license or outright sale, inquiries direct to Knowm.
- [About](https://knowm.ai/about): About Knowm Inc.
- [Newsletter](https://knowm.ai/newsletter): Subscribe for new research notes.
- [RSS feed](https://knowm.ai/rss.xml): Full feed of published posts.

## The chapter series

- [Chapter 1: Transforms will be Assimilated](https://knowm.ai/blog/transforms-will-be-assimilated/)
- [Chapter 2: Before I Knew of the Word Memristor](https://knowm.ai/blog/before-i-knew-the-word-memristor/)
- [Chapter 3: The Thermodynamic Bit](https://knowm.ai/blog/thermodynamic-bit/)
- [Chapter 3b: The Thermodynamic Bit, Derived](https://knowm.ai/blog/thermodynamic-bit-up-close/)
- [Chapter 3c: The Thermodynamic Bit, Built](https://knowm.ai/blog/building-the-kt-bit/)
- [Chapter 4: The Neural Lane](https://knowm.ai/blog/the-neural-lane/): kT-RAM is the genus; the neural lane is one species — addressable kT-bits built from small unit crossbars, selected by an Activation Address Tuple (AAT).
- [Chapter 4b: The Neural Lane Emulator](https://knowm.ai/blog/the-neural-lane-emulator/): kt-ram-neural-core, an open Python emulator of the 2-1 neural lane and its fourteen-instruction set, verified on a single synapse against the Chapter 3b physics.
- [Chapter 5: Thermodynamic Logic Gates](https://knowm.ai/blog/thermodynamic-logic/): Two kT-bits on one neural lane under the FF-RU instruction pair fall into unsupervised attractor states, and those states are the 2-input logic gates — 14 of 16 under a shared encoding, all 16 including XOR under one-hot.
- [Chapter 5b: Maximizing Noise Margin](https://knowm.ai/blog/the-ahah-rule/): Why the Chapter 5 attractors sit where they do — each kT-bit settles its boundary among the nearest opposing patterns, its support vectors, maximizing the noise margin with local updates alone.
- [Chapter 6: Classification and Thermal Sampling on kT-RAM Neural Lanes](https://knowm.ai/blog/classification-and-thermal-sampling-on-kt-ram-neural-lanes/): Supervised classification on neural lanes — one lane per class over an adaptive A2D encoding, a three-case RH/RL/RF instruction routine that matches logistic regression on Iris, and the same lanes read at temperature to draw samples from a chained pair of generator banks.
- [Chapter 6b: The Unsupervised Basis Encoder](https://knowm.ai/blog/the-basis-encoder/): Take the label away and the same routine becomes an online vector quantizer — exclusion and recruitment keep the winner-take-all competition from collapsing, the codebook self-prunes to what the data supports, and sixteen groups plus a decoder and label read-out form a multi-module network on binarized Fashion MNIST whose encode/decode loop, read at temperature, becomes an in-browser generator.
- [Chapter 7: The AAT Codec](https://knowm.ai/blog/the-aat-codec/): The codec that carries whole vectors between real space and AAT space — per-axis binning wastes addresses on correlated data, so a 128-number vector is cut into 64 two-number slots with a small learned codebook each, giving 64 six-bit symbols (48 bytes, a point in A_64^64) that round-trip near 0.98 cosine on held-out Qwen3-0.6B activations; product quantization is Codec A, an additive whole-image variant addresses a garment with 128 atoms, and the codec's second job is kernels — inner products, norms, weighted sums and addition run on the codes through lookup tables instead of multiply-accumulates.

## Reference

- [The kT-bit Catalog](https://knowm.ai/blog/kt-bit-catalog/): A catalog of thermodynamic bits (kT-bits) across nature, scored against three gates.
- [The kT-bit Reading List](https://knowm.ai/blog/kt-bit-reading-list/): The cross-disciplinary lineage behind the thermodynamic bit.
- [A Decade of Knowm SDC Memristor Literature, Part 1](https://knowm.ai/blog/a-decade-of-knowm-sdc-memristor-literature-part-1/): A data-driven survey of a decade of papers that use or cite Knowm's Self-Directed Channel (SDC) memristor and MSS model, with a reproducible full index of 337 papers.
- [A Decade of Knowm SDC Memristor Literature, Part 2](https://knowm.ai/blog/a-decade-of-knowm-sdc-memristor-literature-part-2/): What researchers actually built with the SDC memristor across the decade — neuromorphic synapses, in-memory logic, analog ML accelerators, chaos circuits, hardware security, and the strange ones — sorted by how much work went where.
- [How 374 Papers Build a Signed Synaptic Weight](https://knowm.ai/blog/how-374-papers-build-a-signed-synaptic-weight/): A survey of 374 harvested papers on how the field represents one signed synaptic weight in positive-only memristive devices — one device against a reference column, two read as a difference, or many — with the 1-2 / 2-1 / 2-2 arrangement map and the trade-offs between them.
- [A Taxonomy of Memristor Technologies](https://knowm.ai/blog/a-taxonomy-of-memristor-technologies/): Twenty families of memristive device sorted by switching mechanism rather than material — a computed cladogram over ten coded traits, a dated timeline showing the newest family is a 2017 offshoot of a 1976 branch, a seven-axis star matrix with every number carrying its measurement protocol, why the field's endurance and retention figures mostly cannot be compared, and where Knowm's self-directed channel lands (next to CBRAM), including the axes of Knowm's own scorecard that don't survive the fact-check.
- [ktram-neural-core](https://github.com/knowm/ktram-neural-core): The open-source kT-RAM / neural-lane emulator. MIT License (software only; patents not licensed).