AI Agents That Forget Beat a Single Agent by Up to 39%

A team of AI coding agents that periodically wipes its history and restarts from its best solution scored up to 39% better than a single agent, according to a paper posted to arXiv on 14 October 2026 by researchers listed at UC Berkeley, Stanford and Bespoke Labs. Against four agents working in parallel at the same total token cost, the gain was up to 15%.

The authors, who include Matei Zaharia and Alex Dimakis, write that “the key to scaling discovery further is to forget.”

Business Pill · CONTEXT ENGINEERING

A short explainer on context engineering, what a model sees before it answers. It teaches the general idea only and says nothing about the paper or any organisation in this story.

What the Paper Tested

The paper, “Scaling Discovery Requires Forgetting,” is by Shangyin Tan, Ben Zhou, Lakshya A Agrawal, Jialin Zhang, Yilin Chen, Qiuyang Mang, Donghyun Lee, Dennis Jacob, Koushik Sen, Alex Dimakis and Matei Zaharia. It is a preprint.

It studies autoresearch: a coding agent proposes a solution, a judge scores it, and the agent uses the score to decide what to try next. The authors compare three ways to spend the same budget: one agent that keeps a single session, four agents in parallel that never see each other’s work, and a team of four that shares candidates and forgets.

Forgetting means that at the end of each round every agent discards its session and every earlier candidate, keeping only the best one so far. All four agents then start the next round from that candidate with an empty context. The team forgets three times, which splits its budget into four rounds.

There are four tasks: a heuristic solver for AtCoder Heuristic Contest 014, an optimisation problem from Frontier-CS, Anthropic’s kernel performance challenge on a simulated VLIW SIMD machine, and a five-minute training script for a small language model, nanochat-autoresearch. Every run uses the Pi agent and the DeepSeek V4.1 Flash model, and each configuration ran three times per task, the paper says.

What Was Posted on X

Shangyin Tan, the paper’s first author, introduces it: coding agents running autoresearch for days stop improving early, and the fix the paper tests is forgetting.

Alex Dimakis, a co-author, describes the method as a context-management trick that makes agents forget everything except the best solution so far, and links it to Dimitris Papailiopoulos’s ‘Fork and Flush’ and ‘Idea-basins’ work.

Why the Agents Stall

The authors find that all three set-ups eventually plateau because each agent becomes anchored to its own past attempts. As sessions grow, edits get smaller and reasoning shrinks: before its first reset, the team’s edits fall from about 150 lines to 30 and its reasoning from about 28,000 tokens to 7,000 per turn, according to the paper.

The stall can look like a conclusion. Across the three single-agent runs on the kernel task, 532 of 763 completed turns produced no candidate, consuming 66.2% of the model spend (58.4% to 70.4% per run). In one case, an agent at 1,710 cycles wrote “I’ve reached the practical limit of this machine configuration” and proposed nothing, the authors write.

That was not the limit. The same agent later reached 1,644 cycles, and the team with forgetting reached 1,077 cycles on the same task, the paper says.

What Forgetting Did

The team with forgetting was best on all four tasks. On average it improved on the parallel agents by 10% and on a single agent by 22%, according to the paper.

On AtCoder Heuristic Contest 014, its mean best score was 1,103 thousand against 964 thousand for the parallel agents and 795 thousand for a single agent, 39% above the single agent. On the kernel task, its mean was 1,097 cycles against 1,220 and 1,426, where fewer cycles is better.

On Frontier-CS, its mean score was 80.8 against 70.2 and 64.0. On nanochat-autoresearch the gain was about 2%, the paper says.

In 32 of the 36 forgets, the team found a new best candidate within the next round. The first forget did most of the work: on AtCoder the mean best score rose 11.2% after the first forget and 1.7% after the third, the authors write.

Forgetting Cost Less Than Continuing

In a separate test, the authors took runs that had already plateaued on AtCoder and spent 30 more proposals per agent either continuing the same sessions or forgetting.

For the team of four, continuing gained 1.4% and cost $24.67 per replicate, while forgetting gained 9.2% and cost $15.32. For a single agent, it was 1.6% at $6.66 against 4.5% at $3.44, and for four parallel agents 1.9% at $27.51 against 7.0% at $13.97, according to the paper.

The authors write that forgetting is cheaper because a fresh session carries a shorter context, and that it gained more “at little more than half the cost.”

Where This Sits

The authors list a concurrent paper, Fork-and-Flush by Ziyang Cai and colleagues including Dimitris Papailiopoulos, which forks agents into fresh contexts that inherit the whole workspace. They write that their method “more aggressively” removes the workspace state as well, keeping only the best candidate.

As we read it, the paper adds to a run of results on how long-running agents are organised rather than which model they use. Meta researchers’ 27B agent topped Claude Code on novelty, as we reported. Our explainer on multi-agent systems covers the basic set-ups the paper compares.

The Structural Read

As we read it, the stall the paper measures is behavioural rather than a hard limit: edits shrank from about 150 lines to 30 and reasoning from about 28,000 tokens to 7,000 per turn, and one agent declared a limit at 1,710 cycles that the team later took to 1,077.

As we read it, forgetting is also a cost lever: from the same plateaued state, the team that forgot gained 9.2% at $15.32 per replicate, against 1.4% at $24.67 for the team that kept its sessions, the paper says.

As we read it, the open question is the schedule: the authors write that “a natural next step is to decide when to forget and what to remember,” and in their runs the first forget did most of the work, 11.2% against 1.7% for the third on AtCoder.

Three Implications

TEAMS RUNNING LONG AGENT JOBS On plateaued runs, restarting from the best candidate with an empty context gained more than continuing and cost less, at little more than half the cost, the paper says.

HARNESS BUILDERS Every configuration in the paper is a setting of one component, the reducer, which decides which candidates an agent sees and which session it continues.

ANYONE READING AGENT OUTPUT A single agent that reported reaching the practical limit at 1,710 cycles was not at the limit, according to the paper: the team with forgetting reached 1,077 on the same task.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework’s core claim is that “The surface, not the model, governs autonomy. The agentic surface — the harness — is becoming where value pools in the entire AI stack.”

As we read it, this paper is a controlled case of that claim: the model, the agent and the budget stay fixed, and changing only what the harness lets each agent remember moves the results. The framework predates the paper and does not discuss it or forgetting.

What Is Not Established

The paper is a preprint. Every result uses one model, DeepSeek V4.1 Flash, one agent, Pi, four tasks and three runs per configuration, and on nanochat-autoresearch the gain was about 2%.

The authors write that “a natural next step is to decide when to forget and what to remember”; in the experiments, the team forgets three times per run.

The paper lists research gifts from companies including Accenture, Amazon, AMD, Google, Intel and NVIDIA.

The Bottom Line

A team of four coding agents that discards its history three times and restarts from its best solution was best on all four tasks in this preprint, up to 39% above a single agent and up to 15% above four parallel agents at the same total token cost.

On runs that had already plateaued, forgetting gained 9.2% for the team against 1.4% for continuing, and cost $15.32 against $24.67 per replicate.