Frozen 7B model with external memory cartridges hits 118/128 recalls in solo prototype
Nearby_Indication474 · reddit · 2026-10-07
A Reddit user (ceceli33) has published a solo research prototype that extracts machine-native numerical memory representations from a frozen 7B neural network's internal state and stores them outside the model as "cognitive cartridges." At query time, the input phrase creates a pattern in the network that matches against the whole memory field at once — no filenames or human-readable folder structure needed — and the matching cartridge's numerical state is injected back into the model's latent space, reactivating the memory inside the machine rather than just returning text.
The author reports 118 correct recalls out of 128 test memories (92%) and has released a Zenodo technical disclosure (DOI: 10.5281/zenodo.23198057), an executable demo (MAM128FullDemo.py), raw run logs, and the GitHub repo titan-cognitive-core-v2. The long-term vision: a personal AI accumulating hundreds of millions of such cartridges retrievable by association, mimicking human cue-based memory. This is an independent, non-peer-reviewed prototype — results await verification.
More from Research
- Nature Neuroscience launches special issue on neurotechnology and brain-computer interfaces — plopesresearch · 2026-10-07
- Stanford's Priced Guidance lower-bounds whether LLMs can forecast future research via paid hints — stanfordnlp · 2026-10-07
- Stanford's 'relative scaling laws' show LLM scale isn't a universal equalizer — Diyi_Yang · 2026-10-07
- In AI math, the bottleneck has shifted from finding proofs to verifying them — rohanpaul_ai · 2026-10-07
- Unique Games conjecture rumor confirmed, a major theoretical CS breakthrough — felpix_ · 2026-10-07
- Mathematician littmath: OpenAI Model Proved Bloch's Conjecture, Paper Cites Known-Wrong Work — littmath · 2026-10-07