Beyond RAG: Injecting Tensor Memory Cartridges Into Layer-3 of a Frozen LLM
Nearby_Indication474 · reddit · 2026-10-15
A solo developer proposes a third path beyond RAG and fine-tuning: have a frozen Qwen2.5-7B-Instruct read a fact once, capture its layer-3 residual representation (H3) and attention K/V states from layers 0-3, and package them into a numeric "memory cartridge."
At answer time the source text is never pasted into the prompt; instead the stored H3 representation is injected at the layer-3 boundary, and later transformer layers attend to this numeric memory. Weights stay frozen — no gradients, no LoRA.
The v1.4 addition is cartridge consolidation: three separately written memories can be appended into one consolidated memory while previously written rows stay unchanged. Using three fictional facts forming a multi-hop chain, the model with all three cartridges installed correctly answers three-hop questions; with incomplete memory it sometimes produces unsupported answers (failure cases were explicitly tested). Full technical disclosure and reproducibility docs are on Zenodo (doi:10.5281/zenodo.23381336).
Caveat: this is an independent one-person experiment with limited validation — treat the claims accordingly.
More from Research
- Microsoft's SMD distillation hits 4-step generation that beats multi-step diffusion teachers — microsoft · 2026-10-15
- Colossal Uses AI to Rebuild Extinct Genomes and Track Endangered Species — IgorCarron · 2026-10-15
- Thinking Machines ships Dimuon optimizer for LoRA fine-tuning, beats AdamW in fewer steps — simonguozirui · 2026-10-15
- NVIDIA report: agentic setup lets GPT-6 Astra hit 74.8% on RoboLab-120 robot manipulation — YuXiang_IRVL · 2026-10-15
- Bittensor Subnet 55 completes 10M synthetic CRISPR experiments with 248 competing miners — markjeffrey · 2026-10-15
- Logit lens shows Qwen3-1.7B's refusal to be shut down depends on how many layers you train — Sauers_ · 2026-10-15