---
title: Induced Forgetting
type: vocabulary
url: "https://www.envisioning.com/vocab/induced-forgetting"
summary: "A technique for scaling automated-discovery agents that periodically discards every agent's session and candidate history, restarting from only the best candidate found so far."
year: 2026
generality: 0.40
---

# Induced Forgetting

A technique for scaling automated-discovery agents that periodically discards every agent's session and candidate history, restarting from only the best candidate found so far.
Induced forgetting is a technique for scaling automated-discovery agents in which the search state, including every agent's session history and the full candidate history, is periodically discarded and the search restarted from only the best candidate found so far. Shangyin Tan, Ben Zhou, Lakshya A Agrawal, and eight coauthors, from UC Berkeley, Stanford, and Bespoke Labs, introduced the technique in the October 2026 paper "Scaling Discovery Requires Forgetting" (arXiv:2610.20563). It responds to a plateau the authors observed across three common ways of scaling LLM-agent research: running one agent longer, running agents in parallel, and letting parallel agents communicate and build on each other's candidates. In all three cases, each agent eventually anchors to its own past attempts, makes progressively smaller edits, and stops improving even when given more compute budget.

Rather than letting an agent's accumulated history persist indefinitely, induced forgetting clears every agent's session and the record of past candidates at intervals, keeping only the single best candidate as the new starting point. The authors implement this inside a team of communicating agents, so the team as a whole periodically loses its working memory of prior attempts and failed directions while retaining the best result discovered up to that point. This differs from catastrophic forgetting, where a model unintentionally loses prior knowledge while learning new tasks. Induced forgetting is a scheduled, engineered intervention meant to unblock search, not an unwanted side effect of training.

The authors evaluate induced forgetting on four tasks spanning heuristic programming, GPU kernel optimization, and model training, including an AtCoder heuristic contest problem and a take-home performance-engineering task from Anthropic. A communicating team of agents using induced forgetting produced the best result on all four tasks, improving up to 15 percent over parallel agents without forgetting and up to 39 percent over a single agent at the same compute budget. The paper leaves open how often to forget and exactly what to retain across a reset as questions for future work.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/induced-forgetting)