AkbasCore MAM: frozen 7B LLM with plug-in memory cartridges scores 72/72
Nearby_Indication474 · reddit · 2026-10-09
AkbasCore MAM's first public demo gives a fully frozen Mistral-7B persistent, incremental memory via "cognitive cartridges": tensors appended into the model's internal layers, with no training, retrieval, or source text in the prompt. On a sealed benchmark the append-only memory answers 72/72 — matching the same model reading the full text — while naively stacked cartridges score only 13/72.
Key mechanics:
- A fact passes through the frozen model once; hidden state H6 plus layers 0-6 KV are saved (36 KiB/token, vs 128 KiB for a full KV cache); text is discarded
- Cartridges are written independently, in isolation
- The core discovery: facts bind to each other in middle layers (7-14), so layers 0-6 can stay fully independent; only the upper layers consolidate cartridges (method DC6, bypassing the text entrance)
The author frames this as challenging the Transformer's silent 2017 assumption that everything must share one context window, potentially ending the "one big context window" era. It's an open-source proof of concept, with a scaled version in progress.
Related event: Frozen 7B Models Gain Persistent Memory via "Memory Cartridge" Architecture(2 posts)→
More from Models
- ChatGPT web can actually listen to audio, but the same model can't via API or Codex — Possible_Statement84 · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- TLA+ expert corrects Grok: TLC only exhausts finite state spaces, not all behaviors — tianyin_xu · 2026-10-23
- Open LLM cost calculator shows Arabic tokenization and 6x price gaps across models — Moham-saed · 2026-10-23
- OpenAI blocks manual reset with 13% usage left: 'no reset needed right now' — AccomplishedSugar490 · 2026-10-23
- Meta's 816B Muse Spark Open scores 39, losing to 180B Qwen3.8-Flash-Next at 40 — BTA_Labs · 2026-10-23