Agent memory components: check dreaming and salience notes on named evals, not “Claude has a hippocampus”

Anthropic’s engineering team published “The components of agent memory” on 6 November 2026. Lance Martin frames a concrete problem: agents log work as they go, but knowledge needed across sessions cannot rely on weight updates or a single context window.

The post’s split is dual-track engineering. In session, Claude writes memory notes for what looks salient. In idle time, “dreaming” checks those notes against transcripts, builds index-like wikis, and promotes a small set of items into the next session’s injected context. The test is CL-Bench, Pokémon progress, and Stainless Q&A scores, not a claim that Claude already has a human hippocampus.

Hard facts (https://www.anthropic.com/research/components-of-agent-memory; published Nov 6, 2026; accessed the same calendar day in Asia/Shanghai):

• Definition: agent memory is knowledge from user interactions and assigned tasks that persists across sessions as readable context. The post excludes weight updates, in-context learning, and compaction that replaces older turns with summaries.
• Reference: sleep-time compute from researchers at Letta and UC Berkeley; idle reorganization is called dreaming.
• CL-Bench 1.0: a fresh session per question, with a persistent memory directory. Claude Opus 5 was the first listed model to close the learning loop: most entries included a rule, how the rule was known, and warnings about traps. Opus 5.5 and Sonnet 5.5 also closed the loop; with memory they scored about the same as Opus 5, though their no-memory baselines were slightly higher.
• Pokémon: around step 10,000, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge; Fable 5 wrote nine files in two folders plus a top-level global.md and earned four badges. Opus 5.5 beat the game with a small folder.
• Steering: a “record only failures” rule cut much of the memory gain for Opus 5 and Opus 4.7. Prompting Sonnet 4.6 to read memory before answering and to test its theories nearly doubled its score; Sonnet 5 read memory first on 9 of 10 trials without that prompt, versus less than half for Sonnet 4.6.
• Dreaming correction: a wrong note placed the PokéCenter on the east side of Mt. Moon; without correction Claude dropped off a one-way ledge repeatedly and gave up. After dreaming rewrote the note against the transcript, Claude re-entered the cave and reached Cerulean City.
• Stainless corpus: about 350,000 files, roughly 420M tokens; 25-question eval. Scaling dreaming from 2 to 128 sessions raised graded criteria from 13% to 48%. Wiki plus search rose from 54% to 61% as the wiki grew, versus about 53% with search alone; wiki alone peaked at 48%; wiki plus search saved about 20,000 tokens per question.
• Harvey LAB / merger docs: on multi-document questions, wiki scored 87% versus 73% with search alone; some LAB tasks showed no clear gain; in one case the wiki grew larger than the source files. Contract terms: agents writing one shared table kept 41 of 57 facts (72%); one agent per contract quoting terms word for word kept 57 of 57 (100%).
• Claude Code auto memory loads only the first 200 lines or 25 KB of MEMORY.md into each session.

Evidence strength: this is a first-party engineering note with named evals and file counts, not an independent replication. The Pokémon 413-versus-9 file counts and badge gap, the Stainless 13% to 48% rise, and the contract-terms 72% to 100% fix can all be checked against the post’s figures and prose.

How to check: open the post and verify the CL-Bench loop description, the Pokémon memory-folder comparison, the Mt. Moon correction path, and the Stainless / LAB scores. If later public replications show that “record only failures” does not hurt gains, or that wiki gains do not hold, dial back how necessary dreaming looks.

Steelmanning the post: treating memory as a prediction about what a future session will need fits long-horizon failure modes better than dumping every log. Fable 5 spent less space on on-screen state and more on plans and tested rules, which left more attention for play than Sonnet 3.7’s archive.

Strongest steelman: idle passes against transcripts can fix in-session mistakes (such as a wrong landmark) before the next action; a wiki can turn scattered files into a traversable map, and the post reports a score gap on multi-document questions.

Steelman risks: the evals are mostly internal, and some Pokémon comparisons are described as single runs. The Stainless corpus comes from a company Anthropic acquired, so outsiders cannot audit it directly. The post itself notes that wikis can grow larger than the source, that long injected instruction files reduce adherence, and that dreaming costs compute and may not be immediately available. The hippocampus analogy is rhetoric, not a neuroscience claim.

“Closing the loop” means memory entries carry testable rules and provenance. It does not mean the model has human episodic memory.

For teams shipping agents, the next six months are about whether public benches show reproducible dreaming ablations, whether injected layers like MEMORY.md stay deliberately small, and whether “here is a folder, you judge salience” keeps beating fixed write rules. For readers, the check is file counts, badge or score deltas, and before/after correction paths, not a slogan that AI now “sleeps to remember.”

What to check in this post is the gap between in-session salience notes and idle dreaming on the named CL-Bench, Pokémon, and Stainless/LAB results (for example about 413 files / one badge versus nine files / four badges near step 10,000, and 13% to 48% as dreaming scaled from 2 to 128 sessions). What not to check is whether Claude already has a human hippocampus. If the component split and numbers hold, this is a checkable memory-system design note. If they do not, it remains an engineering story with a metaphor.

【No.2】