Épi agent sets 120 records with persistent memory, beats Codex by 5.4% at 64% of the cost
Arindam_1729 · x · 2026-11-02
The Épi paper's core idea: most agents forget everything when a session ends — Épi doesn't. It keeps persistent research history across model sessions, uses hierarchical memory to resume work, versioned branches to try alternatives and backtrack, an independent review step, and a fixed verifier scoring every output.
Results:
- 120 new records set autonomously: 20 numerical bounds across 12 families, 100 constructions across 8 families
- 84x narrower interval for the n-queens constant; a computer-assisted proof of a case of Tuza's conjecture
- épi-turbo (tuned for coding solvers) beats Codex by 5.4% total score on 30 FrontierOR tasks at 64% of the model cost, same model
Honest caveat: it only works because a fixed verifier exists — without verifiable outputs, memory just helps you repeat mistakes faster. Takeaway for builders: memory + branching + a verifier beats a bigger model.
Related event: Épi agent sets 120 records with persistent memory(2 posts)→
More from coding & agent
- Pinokio Update Brings "Remote AI": Control Local AI Apps From Anywhere — cocktailpeanut · 2026-11-03
- Flock: A New iOS App for Running AI Coding Agents, Built on a t3code Fork — seatedro · 2026-11-03
- WhatIf: 10,000+ steerable LLM agents form a flight simulator for policy — scyrusk · 2026-11-03
- LLMs solving math gets all the hype, but their new infrastructure patterns for reliable software go unnoticed — StewartalsopIII · 2026-11-03
- Vibe coding's 5-year leap: from a GPT-3 React button demo to an Opus 5.5 three.js world — bennash · 2026-11-03
- Stripe's Link Agent Wallet lands in Merge Agent Handler: agents pay without seeing card details — shensi · 2026-11-03