Why Claude Repeats Mistakes: Anthropic's Layered Memory & Verification Fix

The Assistant's Notebook: Not Everything Belongs in One Place

When AI assists long-term—writing articles, organizing research—we want it to remember more so it asks less. Over time its notebook accumulates everything: long-term preferences, intermediate progress, unconfirmed guesses, and obsolete instructions. For example, "write shorter from now on" is a lasting preference; "expand this section" may apply only to one article. Treating the latter as a permanent rule could make every future piece unnecessarily long.

On November 6, Anthropic introduced a layered memory organization method:

Raw Records : what was said and done, kept for later verification.

Experience Notes : distilled from records—confirmed requirements, effective methods.

Knowledge Index : links related notes by topic, showing where to find details.

Startup Reminders : the small, current subset read at the beginning of each task.

Think of a filing cabinet, a notebook, a table of contents, and a sticky note. The cabinet can be huge, but for the job at hand you don't carry the whole cabinet. Importantly, "memory" here means retrievable documents—not retraining the model.

When Notes Are Wrong, Re-reading Doesn't Help

Anthropic used a Pokémon game case. Claude misrecorded a landmark's location, then kept following the wrong note and retraced the incorrect route. Researchers had it cross-check the original session log to see what actually happened. After correcting the note, it navigated correctly.

This is like an assistant writing "Thursday deadline" instead of "Friday." Re-reading the wrong note only reinforces the error; you must pull the original chat or email to discover the mistake. Anthropic calls this inter-task review process dreaming : compare raw records, verify notes, then decide how to fix them.

Therefore, important notes should retain source, date, and scope. "Client dislikes this proposal" is too vague; "Tuesday meeting: client said pricing too high, other parts undecided" enables later lookup. The former might cause abandoning the whole proposal; the latter prompts revisiting pricing.

Raw records can also contain errors. Cross-checking first resolves "did we record it wrong?"; external facts still need documents, data, or fresh confirmation.

Well-Organized Knowledge Bases Can Still Miss Critical Facts

As notes grow, linking them into a wiki-style knowledge base aids retrieval. But organization can introduce errors. In a contract-document test, multiple AI assistants updated the same summary table, overwriting each other and dropping key information.

The fix: each contract got its own page with a dedicated writer, and original clauses were preserved. After the change, retained evaluation facts rose from 41 to 57 (out of 57 required) for that specific task. This does not mean the method guarantees 100% accuracy for all business.

It resembles several people editing a document simultaneously—the final version looks tidy, but earlier contributions may have vanished. Also, bigger knowledge bases aren't always better. Some tasks saw no clear benefit, and one organized set became larger than the source files. If finding things gets harder, check whether too much irrelevant content was copied in.

Thus, when AI says "organized," verify it. Pick a few frequently queried items and ask: Who requested this? Where is the original text? Was it changed later? See if it can retrieve both answer and evidence.

Start Each Task with a Short, Updated Reminder

After storage, the final step: what to read at startup? Technically called "context injection," it's like the assistant glancing at today's sticky note before working.

The reminder must be short and frequently refreshed. Overloading it crowds out space for the current task; inserting a wrong requirement makes the assistant read the same error every time.

For example, Claude Code doesn't load all memory files at once. It reads a limited entry point first, then queries details on demand. Use /memory to view and edit saved notes. Documentation: https://code.claude.com/docs/en/memory

Non-developers can start with an existing project notebook: keep the original, let AI produce an organized copy, review it, then decide whether to adopt.

Prompt for manual organization: "Help me organize these notes without changing the original file. Find duplicates, contradictions, outdated items, or unsupported claims. For each suggested change, tell me: based on which record and what date. Mark uncertain items 'to confirm'—don't guess. Finally give me a concise new version for me to check first."

Anthropic also offers a developer-facing Dreams feature (research preview, access by application, billed by model usage). It generates a separate memory library preserving inputs for human review before adoption. This does not mean regular Claude chat automatically organizes all memory nightly, nor that subscriptions include unlimited free organization. Documentation: https://platform.claude.com/docs/en/managed-agents/dreams

Next time the AI repeats a mistake, beyond another reminder, ask: "What exactly did you record last time? Pull the evidence, let's fix it together."

Research and charts source: https://www.anthropic.com/research/components-of-agent-memory