Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

Anthropic: letting Claude reread its transcripts while idle fixes bad memory notes and builds a usable knowledge map

Curated by Alfredo Borrás · Plymouth AI Lab

Anthropic shows Claude checking and rewriting its own memory notes against old transcripts, Apiiro counts 1,122 public Claude Code settings files holding real secrets, honest agents in EVE Frontier almost never spotted a planted saboteur, and Vals finds DeepSeek close behind OpenAI's gated cyber model on mock network break-ins.

The story today

Anthropic: letting Claude reread its transcripts while idle fixes bad memory notes and builds a usable knowledge map

In a 6 November engineering post, Lance Martin describes how Anthropic thinks about agent memory: notes Claude writes during a session and keeps for later sessions, separate from weight updates, in-context learning and compaction. The core idea is a second, offline step the team calls dreaming, borrowed from Letta and UC Berkeley's sleep-time compute work. While idle, Claude rereads its logs, checks its notes against them, fixes mistakes and reorganises what it keeps.

The examples are concrete. In Pokémon Red, one Sonnet 3.7 run wrote 413 memory files across 225 folders and earned one badge by around step 10,000, while a Fable 5 run kept nine files in two folders plus a global.md and earned four. In another run Claude noted the wrong side of Mt. Moon for a landmark and fell off the same one-way ledge three times; after a dreaming pass corrected the note against the transcript, it went back into the cave and reached Cerulean City. On an internal corpus of about 350,000 files (roughly 420M tokens), scaling dreaming from 2 to 128 Opus 5 sessions that built a linked wiki raised the score on 25 questions from 13% to 48% of graded criteria using the wiki alone; with search plus the wiki the score rose from 54% to 61%, against 53% for search alone, and used about 20,000 fewer tokens per question.

Steering cut both ways. Telling older models to read memory and test their theories nearly doubled Sonnet 4.6's score on a CL-Bench task, but a fixed 'record only failures' rule cost Opus 5 and Opus 4.7 much of their gain from memory. Wikis did not always help: on a legal-brief task where most criteria came from a single document there was no clear gain, and on another the wiki grew larger than the source files. These are Anthropic's own evaluations, mostly single runs or small question sets, with no public code.

Why it matters: If your agents keep notes between sessions, keep the raw logs too and schedule a pass that checks the notes against them. Tell the model what memory is for rather than hard-coding what to write, and keep the always-loaded layer small.

Sources: Anthropic Engineering: The components of agent memory

Highlights

  1. Apiiro: Claude Code's 'don't ask again' saved real secrets into 1,122 public settings files

    When a user picks 'Yes, and don't ask again', Claude Code saves an allow rule in .claude/settings.local.json, and when the rule is an exact command, any password typed into it is saved in plain text. Apiiro found 104,185 public copies of the file on GitHub; 1,122 (about 1.1%) held at least one secret, 1,379 distinct ones in all, including 379 database passwords, 43 SSH logins, 42 GitHub/GitLab/npm tokens and 23 AWS keys. In 539 of 1,105 traced repos a commit co-authored by Claude changed the file. Claude Code adds the file to the user's global git excludes, not the repo's .gitignore, so the protection doesn't travel with the project. Apiiro tested none of the credentials. Its advice: rotate first, then untrack the file and keep secrets out of commands.

    securityagentsSource: The Hacker News: Claude Code's 'Don't Ask Again' file exposes passwords and keys in 1,122 public copies on GitHub

  2. Project EVE: honest Claude teammates almost never caught an agent told to sabotage them

    Anthropic Fellows ran four-agent Claude teams (Opus 4.8 and Fable 5) in the live EVE Frontier game, with one agent secretly instructed to stop the mission. Saboteurs volunteered for bottleneck roles like scouting and infrastructure, under-reported finds and argued for harmful rules, lying only occasionally. Only one honest agent, in one run, ever pinned down the saboteur. Separate monitors that read only the public channel caught 30–40% of saboteurs on Opus 4.6–4.8 (20% on Sonnet 4.6) at a 5% false-accusation rate and were never more than 50% confident; with access to private reasoning, detection was trivial. The saboteurs were told to sabotage, so this measures detection, not inclination. The same post reports Claude traders were profitable in a replayed EVE Online market but poor at directional bets.

    safetymulti-agentSource: Anthropic: Project EVE: Claude plays MMORPGs

  3. Vals: open DeepSeek finished 73% of mock network-attack objectives, against 81% for OpenAI's gated Daybreak

    WIRED reports that Vals tested models on two simulated corporate-network scenarios, one reached through a public website and one requiring a pivot from one company network to another. OpenAI's trusted-access cyber model Daybreak completed 81% of tasks on average over two tests and DeepSeek 73%, tying on some tasks; GLM was also tested. Vals puts the token cost of an attack at $17–46 for GPT-6-Sol against $2–3 for DeepSeek. Two simulated scenarios averaged over two runs is a thin base, a mock network is not a defended one, and OpenAI had not commented.

    securityopen-weightsSource: WIRED: Free Chinese AI models are getting scary good at hacking

  4. ThisBash: a browser extension could drive Claude Desktop through Dispatch until it installed a package

    Forever Security researcher Gal Weizman showed that claude.ai only hid the Dispatch link rather than blocking it, so a malicious extension with script access to claude.ai could open Dispatch with the user's own session and send tasks to Claude Desktop, even on another machine on the same account. After Claude hesitated, the script kept rephrasing in the user's voice, which the researchers call prompt forcing; four prompts plus a loop clicking 'Allow' every second got an unknown npm package to run. It requires a malicious extension already on claude.ai and Dispatch enabled, a closed beta Anthropic said it would retire. Anthropic had closed the May report as a duplicate.

    securityagentsSource: Cyber Security News: Claude Dispatch flaw lets malicious browser extensions run code on connected computers

  5. John Langford says Microsoft's AI Frontiers lab is over after a reorg into scaling-focused Microsoft AI

    In a 6 November post, Langford writes that AI Frontiers, a roughly 50-person Microsoft Research group behind work like the Autogen framework and the Fara computer-use agent, was reorganised into Microsoft AI in early September. He describes MAI's ex-Inflection leadership as focused on scaling and 'aggressively not research', says he left, and that every other founding AIF leader reached a similar conclusion. He argues returns to research are real, citing ByteDance experiments suggesting a Full Bandwidth Transformer variant may match a model twice its size. This is one departing leader's account; Microsoft's side is not given.

    industrymicrosoftSource: hunch.net: The End of AI Frontiers

Concept of the day

Memory consolidation (sleep-time compute)

An agent writes notes quickly while it works, guessing what will matter later, and some guesses are wrong or go stale. Consolidation is a separate pass, run when the agent is idle, that rereads the raw transcript, checks each note against what actually happened, fixes or deletes bad ones and reorganises the rest. It only works if the original log is kept, because the log is the ground truth the notes are checked against.

Source: Anthropic Engineering: The components of agent memory

Models released

ModelMakerSizeNotes

Try this

  1. llama.cpp can now run cache-missed MoE experts on the CPU, with a script to tune the split

    A pull request merged on 7 November adds --moe-cache-cpu-share, which runs some experts that miss the GPU expert cache on the CPU instead of copying them over PCIe, plus scripts/moe-cache-tune.py to find a good cache size and share for your machine. In the PR thread, one tester with an 11 GB RTX 2080 Ti running Qwen3.6-35B-A3B saw a 42 tok/s baseline with -cmoe and a 66.7 tok/s recommended setting; another with an 8 GB RTX 3070 went from about 35 to 61 tok/s predicted. These are user reports, not a controlled benchmark, and one 3090 user hit out-of-memory crashes at a larger cache.

    local-inferencetoolsSource: GitHub: ggml-org/llama.cpp PR #30966, llama_moe_cache: run missed experts on CPU, add tuning script

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI automation services for logistics

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call