The Silicon Hippocampus: Persistent AI Memory via Physical KV Cartridges on SSD

Nearby_Indication474 · reddit · 2026-10-12

The author's team presents AKBASCORE MAM v1.3.0, a "Silicon Hippocampus": a physical memory system built around a frozen model's own internal representations — not saved conversations, not a text database, not fine-tuning.

Core idea: the model forms the memory, SSD stores it, RAM finds it, and the frozen model recalls it.

WRITE: a frozen Qwen2.5-7B-Instruct reads a fact once; lower-layer attention states are captured and remaining K/V states are reconstructed from the layer-3 hidden state, yielding a CUT3 cartridge of 56 bfloat16 tensors across 28 layers, saved as a standalone .pt file with a SHA-256 manifest entry. The published run created 1,453 cartridges (3.97 GB).
Deterministic replay: rebuilding one cartridge from its source fact reproduced all 56 tensors bit-for-bit, with weights unchanged.
READ: given only a question (no source text), question representations are projected into a PCA128 address space; FAISS IVF128 + TOP2 + the locked PEAK50 rule narrow 1,453 → 103 → 16 → one cartridge, which is verified and loaded into the attention cache so the model answers without re-reading the original fact.
Not RAG: text RAG pastes found text into the prompt; MAM restores the model's own stored K/V states directly into attention memory — no vector DB, no SQL, no training.
Limits: 96 questions reuse 48 facts; PEAK50 fixed after TEST626; 2.73 MB per fact; A100 only.

Original post →

The Silicon Hippocampus: Persistent AI Memory via Physical KV Cartridges on SSD

More from coding & agent