Wink Pings

Anthropic's Real Test: The More an Agent Stores, the Faster It Forgets

Anthropic engineers tested the real mechanism behind Claude's cross-session memory. 6 parts, and every conclusion is counterintuitive: the more it stores, the faster it forgets. One prompt to put into Claude Code.

Your agent passes every test, and the issue is still open.

rari's summary is spot on. A better prompt won't fix it, because the model isn't the broken part.

Anthropic engineers ran a set of experiments measuring exactly the thing around the model: cross-session memory. The conclusion is counterintuitive: it isn't about storing more.

An agent that stores everything remembers nothing.

The experiment is broken into 6 parts.

Purpose

Tell Claude what memory is for, not how to use it. The fixed rule 'only record failures' actually cost the strongest model most of the gains.

Small

In the Pokémon task, Sonnet 3.7 stored 413 files and 225 folders and earned 1 badge. Fable 5 stored 9 files and 1 global.md and earned 4.

Store what isn't visible

Strategies, plans, validated rules. Don't store what the screen is already showing each round. Things on the screen don't need memory; memory stores reasoning and conclusions.

Dreaming

Claude rereads notes, checks them against the conversation transcript, and corrects wrong notes. The corrected notes got it through Mt. Moon; the wrong notes left the task stuck forever. If notes aren't maintained, memory goes stale.

wiki

Link pages, grow across sessions. From 2 sessions to 128 sessions, the rubric achievement rate rose from 13% to 48%. Memory isn't a pile of files; it's an interconnected network.

A small index

Each session loads only the first 200 lines or 25KB of MEMORY.md. The rest is read on demand. The index is an entry point, not a warehouse.

image

Two people in the comments added two sharp points. One points at the index: it's the only page loaded every session, and if it's written as a topic heading, it will never trigger. 'auth notes' just sits there; only 'when migration hits login, read auth notes' will open it. The other sums it up more precisely: memory is a working set, not an archive. Save what can change the next decision; keep the rest searchable.

These 6 points aren't isolated prompt tricks. According to rari's framework, they belong to the harness layer—the space where the agent works, including files, tools, sandbox, rules, and memory. rari breaks agent engineering into three layers: harness is the space, loop is retry time, graph is the path the work takes. Three nested layers: the model runs inside the loop, the loop runs inside graph nodes, and the graph runs inside the harness.

6 points, one prompt, put into Claude Code. The full engineering guide is in the quoted tweet.

发布时间: 2026-11-22 18:06