"Knowledge cartridges" for a frozen 72B model: pull the cartridge and the AI honestly says "I don't know"
Nearby_Indication474 · reddit · 2026-10-29
AKBASCORE Silicon Hippocampus 1.7 (TEST815) is an open personal research project that adds "cognitive cartridges" to a frozen 72B model. When the model reads a document, its internal activations (millions of numbers per page) are captured and saved as a sealed file; plugging the cartridge back in restores knowledge of that material without re-reading — no retraining, weights untouched, and each cartridge carries a SHA-256 fingerprint for tamper detection.
The author frames this as a Matrix-style alternative to the Skynet trajectory: human experts (law firms, doctors, engineers) become producers of AI memory — licensing, updating, and revoking their own knowledge cartridges — rather than having it scraped into someone else's model.
Results: given 22 cartridges (2 real papers, 20 decoys), retrieval purely from internal representations put the correct paper in the top-2 for 16/16 questions, with decoys never selected; results held under rephrased questions. A closed-loop reasoning variant that challenges and fuses memories completed the same 16 decisions with 36 reads vs 64 (44% less work). Full engineering reference with 12 figures, open logs and a Colab demo on Zenodo: https://doi.org/10.5281/zenodo.23736317
More from AGI Musings
- Post Claims DeepMind Has Lost Its Lead in Text, Image, Video, Vision and Voice — zacharynado · 2026-10-29
- Raising the Dead with AI: Creating Lifelike Avatars of the Departed — ArtificialOther · 2026-10-29
- Stanford colloquium: choosing insight over overload in the age of AI — jonc101x · 2026-10-29
- Developer ditches 'deep work': with agents, juggling 100 problems beats flow state — zack_overflow · 2026-10-29
- a16z's Andrew Chen jokes AI agents will exhaust all 415 area code numbers — andrewchen · 2026-10-29
- Scott Aaronson on AI Cracking Decades-Old Math Problems: Are Mathematicians Obsolete? — romanyam · 2026-10-29