The Hacker News community has surfaced a new file format specification called .aamt, developed by asamanthinks.com. The project positions itself as the spiritual successor to GGUF, but with a critical twist: it is built specifically for GPU execution rather than CPU inference. The format claims to unify three traditionally separate domains—model weights, agent memory, and world state—into a single, deterministic file container that can be verified, signed, and executed directly on graphics hardware.

The TERA Addressing System and Geometric Memory

At the core of .aamt’s agent memory implementation is a novel addressing scheme called TERA. Every observation or thought from an agent is projected into a four-dimensional unit cube, creating a point (T, E, R, A) that serves as the memory’s physical address. This approach eliminates the need for external vector databases or embedding servers. Instead, the memory is packed into a hexadeca-tree structure with 64-byte nodes, where recall is performed via geometric routing. The system uses Fisher–Rao distance to rank candidates, effectively turning semantic search into a geometry problem that runs on the GPU.

Performance Metrics and Deterministic Integrity

The developers report significant performance gains on Apple Silicon hardware. Using an 8B ternary model compressed to 3.23 GB, the system achieves 41.2 tokens per second sustained decode on an M1 Max. For agent memory, the format claims 317 nanoseconds per route for one million records, with file opening times as low as 0.03 milliseconds. Crucially, the packing process is deterministic; identical inputs always produce identical byte sequences and hashes. This integrity is enforced through FNV-1a-64 hashing and a self-repair mechanism that uses Reed–Solomon error correction and mirrored storage to recover from file corruption.

Self-Repair and Security Features

Beyond simple storage, .aamt introduces a self-repair capability where every memory record is stored twice—once in its original location and once at its reflection in the opposite half of the tree. If one copy is damaged, the system can rebuild it from the mirror. The format also includes robust security layers, supporting AES-256-GCM encryption with a hierarchical key tree. This allows for 'final erasure,' where destroying a specific key permanently renders the associated data unreadable. Additionally, the format supports CRDTs (Conflict-free Replicated Data Types) to ensure that multiple writers can merge changes without conflicts, maintaining hash consistency across different machines.

Key Takeaways

  • .aamt is designed as a GPU-native alternative to GGUF, targeting WebGPU and Metal execution.
  • The format unifies model weights, agent memories, and world states into a single container.
  • Memory recall uses geometric routing via the TERA addressing system, bypassing traditional vector DBs.
  • Built-in self-repair uses Reed–Solomon correction and mirrored storage to recover from data damage.
  • The system is deterministic, ensuring that the same records always produce the same file hashes.

The Bottom Line

If .aamt delivers on its promise of deterministic, GPU-native memory routing, it could finally solve the fragmentation problem that has plagued local LLM agents. It is a bold architectural shift that moves beyond mere inference to true persistent state management.