.aamt – one GPU file format for model weights, agent memory and worlds

.aamt is a new GPU-native file format that stores model weights, agent memory and scene data in a single file, according to its developer BeyondSight, which demonstrated it running live in a browser tab. The format packs an 8B ternary model, BeyondSight GaIaN, into 3.23 GB and sustains 41.2 tok/s decode on an Apple M1 Max, with a 4k-token prompt answering 184× faster on a second pass; a million memory records pack in under half a second into 69,962 nodes, and opening the file takes 0.03 ms. Memories are addressed as points in a TERA 4-cube and recalled by Fisher–Rao distance rather than a vector database, with each record stored twice under Reed–Solomon error correction and a checksum so a single flipped nibble is repaired exactly.

The file format for minds on the GPU. .aamt is what GGUF was for the CPU, built for the GPU. And it holds more than model weights. Watch: a BeyondSight model writes a town, four people walk it and think, and every thought becomes a coordinate, is packed into a real .aamt file, and is recalled by geometry. All of it is computed live in this tab. What you're watching Four steps run for every thought. In the replay, only the model's words come from the recording. Your browser recomputes everything else and checks each content hash against the one recorded. A model builds the scene BeyondSight GaIaN, an 8B ternary model in a 3.23 GB .aamt, writes the town as JSON. The engine checks it, and each place rises as its object closes in the stream. Rejected drafts are kept in the record, not hidden. Thoughts become coordinates Every observation and every thought is projected to TERA: a point T, E, R, A in the unit 4-cube. The point is the memory's address. No embedding server, no vector database. Memories are packed into .aamt Each thought is added to a hexadeca-tree 16 children per node, 64-byte nodes, cells named by Odu words like 0.13.2 and written as a real .aamt file. The same records always give the same bytes and the same hash, on any machine. Recall is geometry Soft routing walks the four most likely root cells. The candidates are ranked by Fisher–Rao distance, and the nearest memories go into the next prompt. The beams you see are that query. <sub i∈{T,E,R,A}</sub | asin √u <sub i</sub − asin √v <sub i</sub | One format. Weights, memories, worlds. Measured on an Apple M1 Max; every figure is in the receipts that ship with the format. int4, ternary and 1-bit weights packed for GPU matrix-vector kernels and read in place. 41.2 tok/s sustained decode on an M1 Max; a 4k-token prompt answers 184× faster the second time. A million records pack in under half a second into 69,962 nodes. Opening the file is 0.03 ms. The memory you watched grow above the town is the same structure at small scale. Places, people and what they remember share one coordinate system and one file. A scene, an agent and a model can be addressed, diffed and shipped the same way. Packing is deterministic, so a hash is an identity: this page re-packs a recorded run and gets the recorded hashes. BeyondSight models ship signed and watermarked. The same C++ core runs the native CLI, Python, Unity and this page. Nothing on this page re-implements the format. The Voxel engine runs .aamt models in the browser and on the desktop. Press 'Run it live' and the thoughts come from InnerSight running on your own graphics card. It repairs itself. And when it can't, it says so. Every memory is stored twice — the record, and a mirror at its reflection in the opposite half of the tree, a different region of the file — and each copy carries Reed–Solomon error correction and a checksum. Damage the file below and read it back through the same verified read every .aamt reader uses. On the right, the same memories without the checksum, so you can see why it is there. Loading libaamt.wasm and the town's memories… Everything else it does A file that survives damage, converges across machines, keeps secrets, and finds things by geometry. Each card names where its number comes from. Self-repair Each record can be stored twice — the second copy at its reflection in the opposite half of the tree — with Reed–Solomon correction and a checksum. One flipped nibble is repaired exactly; a shredded copy is rebuilt from the other; damage two copies cannot settle is refused, never guessed. Audits itself The chirality gate re-routes every record to its reflection, for real, and checks the mirror stored there is byte-identical. Damage is counted and named, not silently served. Many writers, one result Records carry a hybrid logical clock. Merging is a per-key maximum, so two machines that edit apart and sync in any order end with the same records — and so the same file and the same hash. The JavaScript and Python implementations agree rule for rule. Replays exactly A .aamtj journal records every routing decision against the file's content hash, so any lookup can be replayed later and must land in the same cell. Five layers, one file Frozen weights, the world an agent lives in, the private space where it thinks, the memories it has earned, and the code it runs — side by side in one container, each in its own layer. Identity that cannot be confused Every record's identity includes the layer it lives in, so the same bytes are a different identity in each layer — a world page can never pass for a piece of code. Native in every runtime the platform ships. Keys, and erasure that is final Layers can be sealed under a key tree — root, layer, scope. Each mind's sandbox has its own key, so destroying that key erases the sandbox: what was sealed can never be read again. Signed models BeyondSight models ship with a signed catalog and a watermark, so a copy can be checked against what was published. An index built from the math An optional Odu index lists the records in each of the 256 Odu cells, so an exact query reads only its cell. An adaptive Meji tree adds 17–30× for local searches inside a cluster — and about 1× when the question spans everything, which is said rather than hidden. Where, then which The T.E.R.A address finds the right region; an embedding residual picks the one you meant. Alone the address recalls 0.017; together, 0.331. Streams to disk The streaming writer produces exactly the file the in-memory packer would — the same bytes, the same hash — while reading each payload from its source as it goes. A model that remembers what it read The runtime keeps what a model has read: a follow-up at 4,096 tokens answers in 307 ms instead of 30.4 s, a branched conversation starts 55× faster, and the engine tunes itself to the machine it is on. Run it yourself One command: chat with a BeyondSight model and get a benchmark card for your own machine. npx @asamanthinks/aamt