Retention, not loops: why recursive self-improvement keeps starting over
BandAutomatic2009 · reddit · 2026-10-22
A long-form Reddit essay argues recursive self-improvement isn't yet a loop because improvements get discarded. It breaks the loop into three functions—retention, discrimination, promotion—and identifies retention as the binding constraint, since a context window is supplied input, not capability. It weighs four places memory can live (weights, external store, scaffolding, in-computation state), notes retention is also a safety surface with poisoning and drift risks, and announces MemT 1.0, a 293B attempt at memory state inside the model's computation.
More from coding & agent
- mdancho84: The Valuable Ones Architect Systems — Agents, Workflows, Memory, Eval Loops — mdancho84 · 2026-10-22
- AutoResearch: AI Agents That Run the Full Research Loop Autonomously — mdancho84 · 2026-10-22
- Karpathy open-sources AutoResearch: AI agents that run the full research loop autonomously — mdancho84 · 2026-10-22
- Dev uses Claude Opus 5.5 to build a working replica of ILM's Stagecraft volume in Three.js — jasonkneen · 2026-10-22
- Solo founder ships one-command tool to find X's biggest Claude Code and Codex token burners — ThePeterMick · 2026-10-22
- Open-Source Claude Recruiter Finds Ranked Candidates with Personal Emails at $0.20 Each — brandon_galang · 2026-10-22