中
◀ Back to feed
Anthropic

The components of agent memory

Published 2026-11-06 Collected 2026-11-07 08:00 AI news #ai#industry
Full text

The components of agent memory

Published Nov 6, 2026 · Engineering · Anthropic

Søren Kierkegaard wrote that life can only be understood backward, but it must be lived forward. We've encountered a similar bind when designing memory for AI agents.

As humans, we encode memory constantly as we go about our lives. The hippocampus plays a central role in this, capturing experiences in the moment so that memories form quickly and are immediately available. But only some experiences are consolidated into long-term knowledge. Much of that consolidation appears to happen when our brains are offline. As we sleep, and during quiet rest, the hippocampus replays our recent experiences.

Agent memory can use a similar division of labor. Agents capture much of their work in logs or transcripts. 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 same idea can help agents organize their memory by reflecting on logs to fix mistakes or produce new notes and representations. This process is often called dreaming.

We've found that dreaming has been a useful tool for helping us give Claude effective memory. In this post, we share how we've come to think about agent memory and walk through several evaluations we've used to explore agent memory, including the use of dreaming, and what each component contributes to memory organization.

Claude can use dreaming to organize memory. This can correct mistakes, create notes, build new memory representations, or shape what's preemptively loaded into future sessions.

Memory notes

By "agent memory," we mean knowledge that Claude gains from experience—including interactions with users and tasks it's been assigned—and that it retains between sessions, saved as context it can read later. It excludes weight updates, in-context learning within a single conversation, and compaction.

Memory can be a prediction about what the future will need

Memory is, occasionally, a prediction. By writing down a rule, an agent predicts that the pattern it's describing will be relevant again. On Continual Learning Bench (CL-Bench) 1.0, Claude starts each question in a fresh session but can retrieve files from a persistent memory directory. Claude Opus 5 was the first of the tested models to close the learning loop: most entries included a validated rule, earlier answers, and warnings about traps.

Claude has learned to judge salience

We also evaluated Claude playing Pokémon Red with a writable memory folder. Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge; Fable 5 wrote nine files in two folders plus global.md and earned four badges. Fable 5 used memory for strategy and tested rules rather than facts readable from the screen each turn. Opus 5.5 kept a small folder and beat the game.

Steering Claude to write memory

We've found that it helps to tell Claude what memory is for, but not how to use it. On CL-Bench, simply telling Claude a memory folder was available was enough for Opus 4.7 and Opus 5. Setting a rule to "record only failures" weakened the memory boost for stronger models.

Dreaming can correct memory

Dreaming gives Claude a way to test its notes against original transcripts. In Pokémon, Claude incorrectly wrote that the PokéCenter was on the east side of Mt. Moon, concluded "Mt Moon DONE," and repeatedly fell off a one-way ledge. After dreaming corrected the note against the transcript, Claude re-entered the cave and reached Cerulean City. The uncorrected run never got past the mountain.

Dreaming can build new memory representations

With dreaming, Claude can build wiki-style memory representations. On a corpus of ~350,000 internal files (~420M tokens) from Stainless (acquired in May), scaling dreaming from 2 to 128 sessions raised performance on 25 questions from 13% to 48% of graded criteria. Using search plus the wiki raised aggregate scores from 54% to 61% as the wiki grew, versus 53% with search alone, while saving about 20,000 tokens per question.

Dreaming can promote notes up the memory hierarchy

Dreaming can float notes into context injected at the start of every session. Claude Code's auto memory loads only the first 200 lines or 25 KB of MEMORY.md as an index.

Dreaming can focus memory on a specific goal

On a Harvey LAB-style acquisition brief task, changing wiki-building instructions so each contract got its own page with verbatim terms raised graded contract facts from 41/57 (72%) to 57/57 (100%).

Thinking backward, looking forward

Open questions remain: when Claude should create memory in-session versus offline; how to keep memory up to date without continuous checking; and when transcript search plus representations is enough. Dreaming gives Claude a chance to look backward, organize memory, correct mistakes, and maintain representations for future sessions.

Acknowledgements

Written by Lance Martin with editorial guidance from Caitlin Love, Tom Chivers, Rebecca Hiscott, Jake Eaton, and Blake Wilson.

Source
Title
The components of agent memory
Author
Lance Martin
Site
Anthropic
Published
2026-11-06
Collected
2026-11-07 08:00 · 2026-11-07-08-ai.json
Link
https://www.anthropic.com/research/components-of-agent-memory
Open original ↗

Reposted. Will be removed on request. All rights belong to the original author.

More from this batch

08:00batch · 2026-11-0712 items
>

OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation

Bloomberg sources say OpenAI has wired about $11B so far in a round valuing it at $1.25T, potentially totaling ~$20B—well below a prior ~$30…

TechCrunch / Bloomberg
>

中国国安据报策反 AI 高管 Alice Huang,目标刺探黄仁勋与先进芯片情报

The Verge 转纽约时报:MSS 据称招募 Megaspeed 高管 Alice Huang(与黄仁勋无亲属关系)搜集美国 AI 芯片情报,并可能推动先进芯片对华供应。上周已有「刺探黄仁勋」报道,本次新增「已被招募」细节;众议院三委员会要求黄仁勋闭门说明互动。属国家安全舆情…

The Verge / NYT
>

New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration

Anthropic launches dreaming as a research preview on Managed Agents, plus outcomes rubrics and multiagent orchestration in public beta.

Anthropic / Claude
>

Claude Code 被指把用户邮箱塞进 User-Agent;v2.1.317+ 隐私插件默认拦截(issue 仍开放)

简报汇总 GitHub issue #107757:编码 Agent 向第三方站点发请求时 User-Agent 含邮箱;评论称多段 subagent curl 可复现。贡献者称 cc-plugin-privacy 自 2.1.317 默认开启。对企业:查版本与 /plugins…

Models at Work
>

Agent Plugins 1.0:Amazon、Cursor、Microsoft、OpenAI、Vercel 共同推进的可移植 Agent 扩展标准

开放标准:用 plugin.json + skills/ + mcp.json 打包 Skills 与 MCP,一次构建跨 Codex/Cursor/Copilot/VS Code 等加载。谱系明显借鉴 Claude Code 插件布局;Anthropic 未列联合维护方。对企业…

Vercel / AWS / agent-plugins.org
>

Perplexity now has its own always-on AI agent

Perplexity Teammates are always-on agents on dedicated VMs for ongoing projects, priced on the $200/mo Max plan.

The Decoder
>

Beyond code generation: A developer's guide to agentic engineering

GitHub walks through Plan/Build/Verify/Observe agentic workflows with Copilot as commit volume surges and review latency rises.

GitHub
>

AI Gateway adds DeepSeek Search and Fetch tools

Vercel AI Gateway exposes DeepSeek Search and Fetch as model-agnostic tools billed at $0.75 per 1k successful calls each.

Vercel
>

Chat SDK agents can now stream generative UI with json-render

Chat SDK agents stream typed JSON UI specs via json-render into Slack/Teams native cards.

Vercel
>

Speed Up Image Generation with FLUX 3 Image Klein on NVIDIA Blackwell GPUs

NVIDIA shows TensorRT LLM VisualGen optimizations that speed FLUX 3 Image Klein up to ~3.46× T2I / 4.45× I2I vs uncompiled BF16 on B200.

NVIDIA
>

Ripple-65B Preview delivers strong agentic coding results as a fully Apache 2.0 open US model

Ricursive Intelligence open-weights Ripple-65B Preview (dense 65B, Apache 2.0) reports leading open-US TerminalBench scores at ~$32M total c…

The Decoder / HF
>

TAC-CoT:工具增强的隐式 CoT——隐式段穿插显式工具调用,准度最高 +8.8、前向最高约 6.8× 更省

改进什么:显式 CoT 贵且易累积错误;纯隐式 CoT 难插工具调用。怎么做:用紧凑连续 thought token 做推理段,只在工具调用/观测/答案处保留文本,并按查询自适应段数。数学与 QA Agent 基准上相对 Explicit-CoT 提准并大减 #Fwd。KAIST…

arXiv / HF