Experiments with re-writable n-gram tables for LLM persistent memory
Mrinohk · reddit · 2026-09-19
Inspired by how new offload knowledge into n-gram tables, a hobbyist is experimenting with writable n-gram tables plus mid-runtime backprop as a path to persistent LLM memory.
Findings so far: storing memories as question:answer pairs recalls better than raw facts; grafting onto different layers helps but no pattern yet; the main issue is that memory vectors injected into the residual stream can corrupt unrelated outputs when written too strongly. The hypothesis: memories only need to surface once during chain-of-thought, so weaker writes plus attention should suffice.
Testing uses 500 real Q:A pairs from the author's own agent's RAG database (not synthetic), running Qwen3 0.6B on an 8GB RX 6600XT, with sweeps underway on how recall of older facts degrades as memories accumulate.
More from Research
- Insilico's 4B ADMET Model Hits SOTA+ on 7 of 26 Drug Safety Tasks — DeryaTR_ · 2026-09-19
- Dev builds reward-shaping visualizer to compare how reward maps affect GRPO, PPO and TailRL learning — k7agar · 2026-09-19
- 113 Backtests From One Desk: Why Nobody Logs How Many Times the Agent Tried — QuanTradin · 2026-09-19
- Execution Success Isn't Verification: 60/60 Tasks Ran, 0/60 Were Correct — DMAE1133 · 2026-09-19
- Vals AI, backed by Andreessen Horowitz, wants to be the gold standard for AI benchmarks — TechCrunch AI · 2026-09-19
- IR researchers test Jev as a reranker on DL19/DL20: good and cheap — beirmug · 2026-09-19