Anthropic AI Updates: November 7, 2026

1. Anthropic Shows How “Dreaming” Over Transcripts Corrects and Organizes Claude’s Memory, Lifting a Wiki-Based Eval From 13% to 48%

Anthropic published “The components of agent memory,” describing how Claude writes memory notes during sessions and then uses dreaming, an offline pass over its own logs inspired by sleep-time compute from Letta and UC Berkeley, to correct notes, build new representations and choose what is loaded into future sessions. On a Continual Learning Bench 1.0 database task run on Claude Managed Agents, Claude Opus 5 was the first model to close the learning loop by recording verified rules, how it knew they were true and traps to avoid, and Opus 5.5 and Sonnet 5.5 did the same; telling stronger models to “record only failures” removed much of their gain from memory, while explicit prompting nearly doubled Sonnet 4.6’s score. In Pokémon Red, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files plus a global.md dashboard and earned four, and a dreaming pass that checked a wrong note about Mt. Moon against the transcript let Claude get through a section where the uncorrected run never did. Over 128 dreaming sessions on about 350,000 internal Stainless documents (roughly 420M tokens), an Opus 5 wiki raised scores on 25 questions from 13% to 48% of graded criteria, wiki plus search reached 61% versus 53% for search alone while saving about 20,000 tokens per question, and a self-revised wiki recipe for a Harvey LAB contract task kept 57 of 57 graded facts instead of 41. The post also notes limits: one wiki grew larger than its source documents, and promoted notes must stay small because Claude Code’s auto memory loads only the first 200 lines or 25 KB of MEMORY.md. Source

2. Claude Code 2.1.319 Backgrounds MCP Tasks, Adds a Deny-Instead-of-Prompt Setting and a /claude-api latency-optimize Command

Anthropic released Claude Code 2.1.319 on November 6. When an MCP server runs a tool call as a task, Claude Code now moves it to the background immediately and tells Claude the result when it finishes, with CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND to wait inline instead. A new permissions.denyInsteadOfPrompting setting denies any tool call that would otherwise wait on a permission prompt and lets the turn continue, in terminal sessions, subagents and teammates, which suits unattended runs. /claude-api latency-optimize looks for places where Claude API code makes users wait and tries faster settings against the user’s own eval, and CLAUDE_CODE_DISABLE_FEATURE_FLAGS stops feature-flag fetching without disabling telemetry. For the Claude apps gateway, the release adds daily, weekly and monthly spend in /usage for developers with no spend limit, admin.pooled_conflict_rule and admin.pooled_priority_order to decide which pooled group budget pays first, an operator-defined login_notice, JSON log output via CLAUDE_GATEWAY_LOG_FORMAT=json, and start-up warnings for policies affected by a Claude Desktop change scheduled for November 10. It also warns when a session falls back to an ant auth login sign-in that bills a Claude Console organization, and adds an “Earlier in this session” row to /tasks for reopening finished subagents. Source

3. Anthropic Guest Post From Tübingen Researchers Argues AI Modeling Agents Should Return a Distribution of Models, Not One Best Fit

Anthropic published a guest post by AI for Science grantees Jakob Macke, Daniel Gedon and Stefan Wahl of the ELLIS Institute Tübingen and the University of Tübingen on how AI is changing the three steps of scientific modeling: building a model, constraining its parameters and interpreting it. Using the Hodgkin-Huxley neuron model as the running example, they describe simulation-based inference, which trains a neural network on many simulator runs (about 100,000 for Hodgkin-Huxley, millions for larger models) so it can return the full range of parameters consistent with new data in seconds, and note that LLM-based agents can now write candidate models from data and a natural-language description, fit them and revise their equations. Because many differently shaped models can fit the same data equally well, the authors argue agents should report the set of plausible models, whose agreements show what the data supports and whose disagreements point to the next experiment, and they call interpretation the hardest step to hand to AI. Source