Frozen Qwen2.5-72B shows SSD-persistent neural memory with 80/80 bit-exact cross-session reload
Nearby_Indication474 · reddit · 2026-10-25
An independent Reddit research writeup (TEST782, AkbasCore 'Silicon Hippocampus' v1.6) demonstrates an alternative to RAG/text summarization: persisting a frozen Qwen2.5-72B-Instruct's internal numerical states.
- Native tensor fusion: session 1 stores internal states (H3, layer 0–3 K/V) as SHA-256-sealed 'memory cartridges' on SSD; a second session fuses two documents' memories into a single MASTER02 purely via the model's own computation, with no text summaries or weight updates.
- Isolated transfer: sessions run under different Linux users with no file access; GPU memory fully released between sessions — only the SSD-stored numerical memory survives.
- Blind readout: a third session with no access to source docs reconstructs the 80-layer memory with 80/80 bit-exact parity and produces answers referencing both documents.
- Full PDF, 12 figures, and records are on Zenodo; the author notes the semantic synthesis claims have not been peer-reviewed.
More from coding & agent
- 19 latency patterns for faster AI apps: locality, work reduction, concurrency, anticipation — bibryam · 2026-10-25
- 10 open-source repos that turn the web into AI-ready data — Shruti_0810 · 2026-10-25
- popquiz: a free CLI that quizzes you on whether you understand the code AI wrote for you — Honest_Traffic_8613 · 2026-10-25
- Old Android phone + Termux runs 24/7 AI agent, trains PyTorch MNIST to 99.59% — PMinervini · 2026-10-25
- New agent evals field guide draws fire: only 15 of 313 citations predate 2023 — fredahshi · 2026-10-25
- Spotify repositions itself as an AI agent platform with conversational media features — teropa · 2026-10-25