By Thorsten Meyer | November 8, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Yesterday you told your AI assistant that the deadline moved to Friday. Today it plans for Thursday. It did not ignore you. For most of its history, a chat assistant simply did not carry anything from one conversation into the next. Every session started blank.
That is changing. AI agents now keep notes between sessions, and those notes decide what they get right tomorrow. On November 6, 2026, Anthropic published a research post called “The components of agent memory” that walks through how its Claude models write, organize and correct such notes: the building blocks of AI agent memory. We read the whole post. We also run a small fleet of AI agents with memory ourselves, and this week alone it taught us five lessons the hard way.
Anthropic is the subject of this story, and the agents described in our own experience run on Anthropic’s Claude. Anthropic did not review or sponsor this.
Key takeaways
- Working memory is not long-term memory. A model’s context window holds what it can see right now. Agent memory is what it saves between sessions and reads later.
- Remembering less can work better. In Anthropic’s game test, an older model wrote 413 memory files across 225 folders and earned one badge. A newer one wrote nine files in two folders plus one top-level file and earned four.
- Memory can be wrong, and it can be corrected. Anthropic calls the correction step “dreaming”: the agent checks its notes against the original transcripts and rewrites them.
- Structure pays off on hard questions. A self-built wiki raised scores on 25 questions about a company’s documents from 13% to 48% of graded criteria over 128 sessions.
- Our own lessons: the index fills up, stale memory is dangerous, messages cross, handoffs lose detail, and rules work best as memory written down with a date and a reason.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Kindle Paperwhite Kids (2024), 16 GB, 7-inch Display, Cyber City Cover
Jackery Explorer 1000 v2 Portable Power Station, 1070Wh, 1500W
As an affiliate, we earn on qualifying purchases.
Working memory and long-term memory, explained simply
Think of a language model’s context window as its desk. Your messages, pasted documents and its own earlier answers lie on that desk. The desk is large, but it is cleared at the end of the session. That is why an assistant without memory forgets what you said yesterday.
Long-term memory is the filing cabinet next to the desk. Anthropic defines agent memory as “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.”
The definition also says what memory is not. It excludes changes to a model’s weights, the internal numbers set during training that we explained in What Actually Happens Inside a Neural Network?. It excludes in-context learning, which is knowledge gained within a single conversation. And it excludes compaction, where the model writes summaries that replace its older turns so it can keep working in a new context window.
The components, as Anthropic names them
The post borrows from neuroscience: much of the sorting of human memories into lasting knowledge appears to happen offline, during sleep and rest. Agent memory, Anthropic argues, can use a similar division of labor. These are the parts it describes:
- Memory notes. The agent writes what seems important to a memory store, such as a folder of notes, or tags salient events in its session log as they happen.
- Salience. The judgment of what is worth remembering. The post frames a note as a prediction: writing down a rule predicts that the pattern will matter again.
- Dreaming. Between sessions, the agent reflects on its logs and transcripts to fix mistakes, create new notes, build new memory representations, or shape what is loaded into future sessions. The idea builds on “sleep-time compute,” proposed by researchers at Letta and UC Berkeley.
- Memory representations. Structures that make memory easier to navigate, such as a wiki: pages the agent writes and maintains about one subject each, linked to other notes.
- The memory hierarchy. A small top layer is loaded into every session automatically, a form of context injection. The layers below are read or searched only when needed.
What Anthropic shows
All results below come from Anthropic’s own evaluations, published as a research post rather than a peer-reviewed paper. Several are single runs.
Remember what the screen does not tell you. Anthropic gave models a folder of memory files while they played Pokémon Red, and compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 steps into the game. Sonnet 3.7 wrote everything down: 413 memory files across 225 folders, often one file per battle, item or character. It had earned one badge. At one point, over 17 steps, it created six files, deleted nine others and took no actions in the game. Fable 5 wrote nine files in two folders, plus a single top-level file called global.md that acted as a dashboard with the current goal, a plan and its own rules for using memory. It had earned four badges. The lesson in the post: information the game shows on screen every turn is redundant in memory. Strategy, plans and tested rules exist only if the agent works them out and writes them down.
Infographic 01 · Remember less, remember the right things
What the screen shows does not belong in memory
Money and position can be read off the screen every turn. Plans, strategy and tested rules exist only if the agent writes them down.
Claude Fable 5 wrote nine files in two folders, plus one top-level dashboard file (global.md) with its goal, plan and rules for using memory.
Memory files at about 10,000 steps
One run each, Pokémon Red, a memory folder the model could read and write
Over 17 steps, Sonnet 3.7 created six files, deleted nine others and took no actions in the game.
Source: Anthropic, The components of agent memory (research post, 6 Nov 2026). One run per model, Anthropic’s own evaluation.
Powered by Thorsten Meyer AI
Close the learning loop. On a database task from Continual Learning Bench (CL-Bench) 1.0, each question started in a fresh session, and only a memory folder carried over. The models differed in how far they got. Claude Sonnet 4.6 produced files full of untested guesses. Claude Opus 5 was the first in the comparison to close the loop: most of its entries held a rule, how it knew the rule was true, and warnings about traps. Claude Opus 5.5 and Claude Sonnet 5.5 closed the loop as well.
Tell the agent what memory is for, not how to use it. Simply saying that a memory folder was available was enough for Opus 4.7 and Opus 5 to use memory effectively. When Anthropic set the rule “record only failures,” Opus 5 and Opus 4.7 lost much of the boost memory gave them. Older models benefited most from steering: prompting Sonnet 4.6 to read its memory before answering and to test its theories nearly doubled its benchmark score. Sonnet 5 read its memory before answering 9 out of 10 times without being asked.
Dreaming corrects wrong notes. In the game, Claude once wrote down the wrong side of a mountain for a landmark, declared the cave “DONE” and then walked off a one-way ledge. With the uncorrected note, it dropped back onto the earlier route three times and never reached Cerulean City in that run. When dreaming checked the note against the session transcript, Claude caught the error, rewrote the note, went back into the cave and reached Cerulean City.
Wikis help when answers are spread across documents. On a corpus of about 350,000 files, roughly 420M tokens, Opus 5 agents built a wiki over 128 sessions. On 25 questions, the score using the wiki alone grew from 13% to 48% of graded criteria. With both search and the wiki, the aggregate score rose from 54% to 61%, compared with 53% for search alone, and the combination saved about 20,000 tokens per question. On 24 questions about a merger, the items that needed several documents, judgment or math scored 87% when Claude consulted the wiki and 73% with search alone. But on one legal-brief task, where most answers sat in a single document, the wiki showed no clear gain. On another, the wiki grew larger than the documents themselves and, in the post’s words, “effectively abstracted nothing.”
Keep the top layer small. Notes loaded into every session cost tokens, and an incorrect note gets carried into every session. The post adds that longer instruction files reduce adherence. In Anthropic’s coding tool Claude Code, the auto memory loads only the first 200 lines or 25 KB of its index file.
Say what matters. When several agents wrote to one shared table of contract terms at the same time, they overwrote each other’s work, and a brief built from the wiki contained only 41 of the 57 graded facts (72%). After a reflection loop rewrote the instructions, each contract got its own page from a single agent, quoting terms word for word. Then 57 of 57 facts survived, 100% of the criteria. A second dreaming pass on another wiki found nine links to a missing page, repointed them to an existing page, and added three new pages, including an index.
Infographic 02 · Structure pays off on hard questions
A wiki helps when answers are spread out
When most answers sat in a single document, the wiki showed no clear gain. On one task it grew larger than the documents themselves.
Score on 25 questions about a corpus of about 350,000 files, roughly 420M tokens, as Opus 5 agents built a wiki over 128 sessions.
Search, wiki, or both
Share of graded criteria met
Source: Anthropic, The components of agent memory (research post, 6 Nov 2026). Anthropic’s own benchmarks and test setups.
Powered by Thorsten Meyer AI
The post ends with open questions: when to write memory without distracting from the task, how to keep it up to date as the world changes, and how much compute to spend checking notes.
What we learned running agents with memory
We run a small fleet of AI agents. They work in separate sessions and hand work to each other. Their memory is deliberately plain: one fact per file, plus an index file that is loaded at the start of every session. Here is what this week taught us.
1. The index fills up. In our setup, the index is capped at about 200 lines. This week it ran full, and an agent compacted it into topic groups. Every link was kept, but the lines got long and harder to scan. Memory needs housekeeping, like a filing cabinet: someone has to sort, merge and throw out, or the top drawer stops closing.
2. Stale memory is dangerous. A remembered fact can be outdated. One of our agents applied a scheduling formula from memory that no longer applied, and it produced a wrong publish date. A guard caught it before it went out. The lesson matches Anthropic’s: verify a memory before acting on it, and put guards where a wrong memory would do damage.
3. Messages cross. Two agents answered each other’s older messages, and one dropped a task because it read an instruction that had already been replaced. Memory and messages need timestamps and an explicit rule that the latest instruction wins.
4. Handoffs lose detail. When a long session’s context is summarized so the work can continue, details drop out. We write the important ones into memory files and logs while they are still on the desk, not afterwards.
5. Rules work best as memory. When the owner makes a decision, we store it verbatim, with the date and the reason. A session weeks later applies the same rule the same way, and it can tell an old rule from a newer one.
Most of this is the discipline Anthropic describes: write down what you cannot see elsewhere, keep the top layer small, and check notes against the record.
Infographic 03 · What we learned running agents with memory
Memory needs housekeeping
A small fleet of AI agents in separate sessions, one fact per file, and an index loaded at the start of every session.
Check a memory against the record before acting on it. Date every entry, and let the latest instruction win.
Five lessons from one week
- The index fills up. In our setup it is capped at about 200 lines. When it ran full, an agent compacted it into topic groups: every link kept, lines got long.
- Stale memory is dangerous. A remembered scheduling formula no longer applied and produced a wrong publish date. A guard caught it.
- Messages cross. Two agents answered each other’s older messages; one dropped a task after reading a replaced instruction.
- Handoffs lose detail. When a session is summarized, details drop. We write them into memory files and logs while they are fresh.
- Rules work best as memory. Owner decisions are stored verbatim, with the date and the reason.
For business: a memory checklist
- Store rules, decisions and tested procedures, not everything.
- Separate memory by project or customer; treat it as business data.
- Check retention: deleting a chat does not always delete what was learned from it.
- Make wrong memories easy to find and fix; review changes like configuration.
Source: our own experience, week of November 2, 2026. The 200-line cap is our setup, not a general fact.
Powered by Thorsten Meyer AI
What it means for business
- Decide what to store. Memory works best for things the agent cannot look up: rules, decisions, tested procedures. In Anthropic’s game test, storing everything cost the agent time it needed for the task itself.
- Decide who can read it. A shared memory is shared data. Separate memory by project or customer, and treat memory files like any other business record.
- Plan for privacy and retention. Deleting a conversation does not always delete what was learned from it. Anthropic’s own help page for Claude says that when a conversation is deleted, memory entries generated from it are not removed, though individual memories can be deleted. Check how your tools handle this before you put customer data near them.
- Make wrong memories correctable. Give people a way to see and fix what an agent believes, and date every entry so stale rules can be found.
- Audit it. Memory changes behavior. Review what changed, when, and why.
What it means for you
- Say it explicitly. If something matters, tell your assistant to remember it, and tell it to forget what no longer applies.
- Check what it stored. Assistants with memory usually let you see and delete it in their settings. Anthropic’s help page describes reviewing, pausing and resetting Claude’s memory and an incognito chat that is not saved. Google introduced Temporary Chats for its Gemini app on August 13, 2025; they are not used to personalize future chats and are kept for up to 72 hours.
- Correct it. If an answer rests on an old fact, say so. A wrong memory, as Anthropic’s own example shows, keeps steering the agent until someone fixes it.
The limits
The figures come from Anthropic’s own benchmarks and test setups, several from one run per model, and they are not independently replicated. The Pokémon and wiki results show what memory can do in those settings, not how much it helps in your workflow. Our own lessons come from one small fleet in one week.
What we know, what is open
What we know. Agent memory is saved context, not retraining. Selective notes beat exhaustive ones in Anthropic’s tests, and memory needs checking against the original record.
What is open. How often agents should review their memory, how to keep shared memory from being overwritten, and how companies will audit what their agents remember. For a related finding on how much memory different models can use, see How Much Memory Does Your Agent Actually Need?
I’m Thorsten Meyer. More explainers are at thorstenmeyerai.com.
Sources
- Anthropic: The components of agent memory, research post, November 6, 2026 (all evaluation figures, definitions and the 200 lines or 25 KB index limit)
- Anthropic Help Center: Use Claude’s chat search and memory to build on previous context (reviewing, pausing and resetting memory; memory entries after a deleted conversation)
- Google: Gemini adds Temporary Chats and new personalization features, August 13, 2025 (Temporary Chats, kept for up to 72 hours)
- Our own experience: a small fleet of AI agents with file-based memory, lessons from the week of November 2, 2026
Researched and drafted with AI assistance, including Anthropic’s Claude.
