Most AI audio tools are built around a single promise: give them a prompt, get back something polished. Engram, a new instrument from music startup Thoughtful Things, works from the opposite premise. It treats AI failure not as a bug to be fixed, but as the raw material of a new kind of music.
The device is a sampler and groovebox. It takes incoming audio, runs it through a locally hosted AI model, and produces sounds that are distorted, glitchy, and deliberately strange. The goal is not to approximate what a professional studio would output. The goal is to find what lives at the edges of what a small AI model can do, and then push past those edges.
A Tiny Model, Trained With Intention
Engram does not connect to the internet. The AI running inside it is compact by design, trained in-house by Thoughtful Things on audio datasets licensed for commercial use, specifically material under CC-BY or similar open licenses. The company has stated explicitly that it has not trained its models on non-commercial, pirated, or otherwise unlicensed data, and that it never will.
This is worth pausing on. The question of what data AI models are trained on has become one of the central legal and ethical disputes in the creative industries. By committing publicly to licensed-only training data, Thoughtful Things is staking out a position that many larger AI audio companies have been reluctant to take with the same clarity. Whether that commitment holds as the company scales is a separate question, but the declaration itself is notable in a space where transparency has often been scarce.
The choice to run the model locally, rather than through a cloud connection, also carries meaning. It makes Engram a self-contained instrument, more like a synthesizer than a software subscription. The user owns the device and its behavior. There is no server dependency, no usage policy that can change overnight.
Circuit Bending for the AI Era
Founder Evan King describes Engram as a “field recorder for latent space.” The phrase is worth unpacking. Latent space, in machine learning, refers to the internal mathematical representation a model builds of the data it has learned. When a model generates output, it is navigating that space. Engram’s premise is that navigating it badly, or pushing it into territory it was not designed to handle, produces something interesting rather than something broken.
In the campaign video, King demonstrates this by speaking the word “piano” into the device. What comes back is not a piano sound. It is something vaguely piano-like, glitched and transformed, recognizable only in outline. That gap between input and output is the instrument.
The reference point here is circuit bending, a practice that emerged in experimental music decades ago. Circuit benders physically modify electronic devices, toys, and instruments by rewiring their circuits in ways the manufacturers never intended, producing unpredictable sounds as a result. Engram applies that same logic to AI: the model can be tweaked, pushed, and effectively broken, and the results are treated as musical material rather than errors.
Thoughtful Things also plans to open up Engram’s firmware, allowing users to modify it or load their own custom models. This positions the device not just as an instrument but as a platform, one that could evolve through community experimentation in ways the company itself cannot fully predict.
Engram is launching through Kickstarter at $675, described as a 30 percent discount from an expected retail price in the range of $850 to $900.
What This Reveals About AI and Creative Tools
The broader significance of Engram is not really about music. It is about a shift in how AI tools are being conceived for creative work.
The dominant model in AI creativity tools has been one of smoothing: reduce friction, eliminate imperfection, deliver output that sounds or looks finished. That model serves a certain kind of user and a certain kind of goal. But it also narrows the space of what is possible. When every output is optimized toward a target, the range of outputs converges.
Engram points in a different direction. It treats the AI model as a material with its own properties, including its failure modes, and asks what happens when a human musician works with those properties rather than against them. This is not AI replacing human creativity. It is a human using AI the way a sculptor uses stone: working with the grain, finding what the material wants to do, and shaping that into something intentional.
The instrument also raises a question that will become more pressing as AI tools proliferate: who controls the model, and on what terms? Engram’s local operation, open firmware, and transparent training data policy represent one answer. It is an answer that prioritizes user autonomy over platform convenience.
In Short
Engram is a music instrument that runs a small, locally hosted AI model and uses that model’s glitches and distortions as its primary creative output. Built by Thoughtful Things, it draws on the tradition of circuit bending to treat AI failure as a feature rather than a flaw. Its training data policy, open firmware, and offline operation distinguish it from cloud-dependent AI audio tools. The deeper point is this: not every AI creative tool needs to optimize for polish. Some of the most interesting creative territory may lie precisely where the model breaks down.
Based on reporting from The Verge.