The Components of Agent Memory
Comments on “The Components of Agent Memory”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Anthropic describes "dreaming," a background pass where Claude reviews session transcripts to correct, reorganize and promote its memory notes. In the company's own tests, scaling dreaming from two to 128 sessions on a 350,000-file corpus raised wiki-based scores from 13% to 48%. On merger questions, the wiki scored 87% against 73% for search alone. Gains were uneven: single-document tasks showed none, and one wiki grew larger than its sources. Most figures come from Anthropic's internal evaluations and are not independently sourced.
Why it ranked
No. 8 of 107 with a score of 0.68. The panel rated it highest on new information (0.87) and reader interest (0.79), lowest on significance (0.65). Placed No. 8, No. 14 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- Detailed, practitioner-relevant engineering account of agent memory design with several concrete evaluations; claims are mostly internal-benchmark based and somewhat promotional in framing, so substantiation is moderate.
- AI judgeJudge 3gpt-5.6-luna
- A detailed, evidence-backed engineering account of agent memory, dreaming, salience, wikis, and evaluations, with concrete results and limitations. It is highly relevant to AI builders, though based largely on proprietary Claude experiments.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.68total under rubric v2.3
- Significance0.65
- Reader interest0.79
- Audience breadth0.68
- Practical relevance0.77
- Substantiated0.77
- Technical depth0.76
- New information0.87
- Counts against: Promotional0.18
- Counts against: Overclaim0.19
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.68 | 0.50 | 0.78 | 0.65 | 0.28 | +0.24 | +0.157 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.65 | 0.80 | 0.91 | 0.79 | 0.26 | +0.26 | +0.205 |
| Audience breadthHow many people who build technology does it concern? | 0.57 | 0.67 | 0.80 | 0.68 | 0.23 | +0.13 | +0.089 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.72 | 0.75 | 0.83 | 0.77 | 0.10 | +0.22 | +0.169 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.85 | 0.60 | 0.86 | 0.77 | 0.26 | +0.05 | +0.039 |
| Technical depthHow technically substantive is it? | 0.79 | 0.70 | 0.80 | 0.76 | 0.10 | +0.05 | +0.038 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.88 | 0.80 | 0.93 | 0.87 | 0.13 | +0.05 | +0.044 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.12 | 0.25 | 0.18 | 0.18 | 0.13 | −0.19 | −0.035 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.15 | 0.30 | 0.12 | 0.19 | 0.18 | −0.11 | −0.021 |
| Sum of contributions | +0.683 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.683 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-07/8/"><img src="https://rubric.hrih.si/badge/2026-11-07/8.svg" alt="Ranked #8 on Rubric, 7 Nov 2026" width="231" height="28"></a>Claim check
Claim check: 2 matches source · 5 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceGeneralization
Researchers at Letta and UC Berkeley proposed sleep-time compute as a way for models to use idle time between interactions to reorganize their existing context.
The cited source says the same thing.
The article says “Researchers at Letta and UC Berkeley proposed sleep-time compute as a way for models to use idle time between interactions to reorganize their existing context”
The source says “We refer to such a process, as sleep-time compute: where inference is done between interactions with the model while it would otherwise be idle in sleep-time”
Source cited by the article: arxiv.org/abs/2504.13171
No source givenNumber
Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Fable 5 wrote nine files in two folders plus global.md and earned four badges, about 10,000 steps into Pokémon Red.
The article gives no link or citation for it.
The article says “Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge”
The source says No source to compare with.
No source givenNumber
Prompting Sonnet 4.6 to read its memory files before answering and to test its theories nearly doubled its CL-Bench score.
The article gives no link or citation for it.
The article says “prompting Claude to read its memory files before answering—and then to explicitly test its theories—nearly doubled its benchmark score”
The source says No source to compare with.
Couldn’t checkFirst, only or fastest
Opus 5 was the first of the tested models to close the learning loop on the CL-Bench 1.0 task.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Claude Opus 5 was the first of these models to close the loop”
The source says No passage could be compared.
Source cited by the article: continual-learning-bench.com/news/cl-bench-1-0
No source givenCausal conclusion
With the corrected memory note after dreaming, Claude went straight back into Mt. Moon and reached Cerulean City, whereas the run with the uncorrected note never reached Cerulean City.
The article gives no link or citation for it.
The article says “With the corrected memory note, Claude went straight back into Mt. Moon and eventually reached Cerulean City”
The source says No source to compare with.
No source givenNumber
Scaling dreaming from two sessions to 128 raised performance on the 25 questions about the Stainless corpus from 13% to 48% of graded criteria.
The article gives no link or citation for it.
The article says “As we scaled dreaming from two sessions to 128, performance on the 25 questions grew from 13% to 48% of graded criteria”
The source says No source to compare with.
No source givenBenchmark result
On merger questions needing cross-document search, judgment or math, Claude scored 87% using the wiki versus 73% using search alone.
The article gives no link or citation for it.
The article says “Claude scored 87% when it consulted the wiki, compared with 73% when it used search alone”
The source says No source to compare with.
Matches sourceNumber
Claude Code's auto memory loads only the first 200 lines or 25 KB of MEMORY.md into each session.
The cited source says the same thing.
The article says “It loads only the first 200 lines or 25 KB of MEMORY.md”
The source says “The first 200 lines of MEMORY.md, or the first 25KB, whichever comes first, are loaded at the start of every conversation”
Source cited by the article: code.claude.com/docs/en/memory
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 7 pts, not on front page
Recorded for context only, checked 7 Nov 2026, 05:47 UTC. It played no part in the ranking.