Claude Memory 2026 — Categorized Entries, from App to API
- See exactly what changed in July 2026 — categorized entries replace the daily summary in the app, and a new beta header rewires the API
- Learn how to view, edit, and delete individual entries in Settings without wiping everything
- Understand the four behavior changes the agent-memory-2026-07-22 header brings to memory_store endpoints, and why the old cursors stop working
- Meet the memory_store.* webhooks and Dreams — the parts nobody documents
- Apply five practical patterns for teaching Claude what to remember (and what to forget)
For a year, "Claude's memory" meant a single evolving paragraph the model rewrote each day — good for a bio, bad for anything with structure. Between June and July 2026 Anthropic replaced that with an entry-based model: dozens of small, categorized notes Claude reads and writes in real time. Same idea powers the app and the Managed Agents memory API. This page connects them so you understand both surfaces from one mental model.
The mental model that unifies both sides
Both sides work the same way now: many small entries at named locations, each rewritable in isolation. The old model (one blob per user, rewritten wholesale) is gone in the app and has never been the API surface. Once you internalize "entries at paths, listed by prefix, edited one at a time", both surfaces stop feeling like separate features.
Part 1 — In the app: categorized entries
What actually changed, and when
- A new Settings › Reflect surface starts showing topics you spent time on, your most active day, peak hour, and observations about how you work with Claude. Beta on Free, Pro, and Max (web + Claude Desktop). Requires memory enabled.
- Memory becomes a set of individual, categorized entries that Claude reads and updates during conversations, instead of one rolling daily summary. Announced in the Claude Apps release notes.
- Third-party coverage places availability across Free, Pro, Max, Team, and Enterprise, but plan-by-plan wording shifts between Anthropic's help center and news posts. When it matters (compliance, audits), verify in-product for your plan.
Why the shape change matters
The old daily summary had a single failure mode: unrelated context leaked into each other. Fix a fact about your work project and Claude might overwrite something about your writing style. Small, categorized entries fix that structurally — each edit is scoped to one lane.
Three consequences most users miss:
- Retrieval is now targeted, not associative. Claude can pull "your communication preferences" without dragging in the rest of your context. Prompts that used to need "remember I prefer X, not Y" to break through the summary now hit an existing entry directly.
- You can prune surgically. Delete a single wrong entry in Settings instead of wiping everything and re-teaching Claude who you are. The "burn it all down" tax on correcting one mistake is gone.
- Categories act as an invisible schema. When Claude decides where an entry belongs (e.g. "communication style" vs. "current project"), it's implicitly choosing whether that fact will be recalled in future personal chats, future work chats, or both. Naming things well up front nudges better recall later.
View and manage your own entries
- You'll see entries grouped by category (professional context, communication preferences, ongoing projects, and similar buckets).
- Each entry has its own controls. Fix a wrong fact without touching the rest — the whole reason for the redesign.
- Ask Claude to remember, forget, or update specific things ("remember I use Bun, not npm") and it will update the matching entry live.
- An incognito chat is excluded from memory writes and chat history. Use it for anything you don't want persisted — sensitive drafts, one-time explorations, credentials you'd rather not have summarized.
- Settings › Reflect surfaces what Claude has been talking to you about. Skim it to catch memory drift — if a topic you no longer care about keeps showing up, prune the entries.
- Memory is best-effort recall, not authoritative storage. Never treat it as a source of truth for anything you'd audit (contracts, credentials, medication, legal facts). Restate hard requirements in the prompt or in a Project's instructions.
Five patterns that actually work
1. Set standing preferences explicitly, once
Please remember these as standing preferences for our chats: - Language: Italian for prose, English for code and error strings. - Units: metric. - Coding style: TypeScript strict, no default exports. - Voice: direct, no filler, no closing "let me know if…". Store each as its own entry in the right category.
2. Correct a single fact without a wipe
Update the memory entry that says I use Vercel for hosting — I moved everything to Cloudflare Workers last month. Leave every other project fact alone.
3. Deliberately forget
Delete any memory entries about the "Project Nightingale" launch — that work is finished and I don't want it surfacing in future chats.
4. Audit what Claude thinks it knows
Before we start, list every memory entry you have about me, grouped by category. Flag anything that looks outdated so I can decide whether to keep or delete it.
5. Sandbox exploratory chats
(Open the chat in Incognito.) Don't try to remember anything from this conversation — I'm just exploring a sensitive draft I don't want stored.
Part 2 — In the API: memory_stores and the new beta header
The Managed Agents platform exposes memory as memory stores — collections of memories addressed by path, listed with a path_prefix, and paged with cursors. On July 2, 2026 Anthropic shipped agent-memory-2026-07-22, a beta header that changes the semantics enough to break naive migrations. On July 22, 2026 the older managed-agents-2026-04-01 header adopted the same list behavior — so the "new" rules are now the only rules on memory-store endpoints.
The four behavior changes you must know
| # | What changed | Old behavior | New behavior under agent-memory-2026-07-22 |
|---|---|---|---|
| 1 | List order | Client-controlled via order_by / order | Stable, server-defined order. order_by and order are ignored — sorting client-side if you need it |
| 2 | depth parameter | Free-form integer | Accepts only 0, 1, or omitted. Any other value returns 400 |
| 3 | path_prefix matching | Substring match | Must end with /, matches whole path segments only |
| 4 | Pagination cursors | Interchangeable | Cursors issued without the new header aren't valid with it. Restart from page 1 when you adopt it |
Two more gotchas:
- Header collision. On memory-store endpoints,
agent-memory-2026-07-22replacesmanaged-agents-2026-04-01. Sending both returns 400. Migrate the memory-store calls only; leave the other managed-agents endpoints on the original header. - SDKs already switched. Python 0.116.0, TypeScript 0.110.0, Go 1.56.0, Java 2.48.0, Ruby 1.55.0, PHP 0.36.0, C# 12.35.0, and the CLI 1.16.0 all send
agent-memory-2026-07-22on memory-store calls automatically. If your code passesbetasexplicitly, replacemanaged-agents-2026-04-01there — don't add both.
Migrate a list-memories call, minimally
- Upgrade to at least Python 0.116.0 / TS 0.110.0 / Go 1.56.0 / Java 2.48.0 / Ruby 1.55.0 / PHP 0.36.0 / C# 12.35.0 (CLI 1.16.0). The header switch happens for you on memory-store calls.
- Remove order_by / order — the server ignores them. If your code sorts on the returned list order, move the sort client-side.
- Replace any depth other than 0 or 1 with the value that best matches your intent, or omit the parameter.
- Ensure every path_prefix ends with a slash. Remove any substring-style prefixes that relied on partial-segment matches — they will silently return fewer results now.
- On the first request after upgrade, don't reuse a stored cursor. Start from page 1 and re-derive the next cursor from the new response.
- If you pass betas manually, replace managed-agents-2026-04-01 on memory-store calls with agent-memory-2026-07-22 — never send both.
Webhooks: react to memory-store lifecycle without polling
Shipped July 22, 2026 alongside the header default: three memory_store.* webhook event types now cover the memory store's lifecycle, and four environment.* events cover the sandbox environment. This is the piece most integrators miss — you can now react to a memory store being created, updated, or removed by an agent run instead of polling GET /v1/memory_stores on a timer. Subscribe on the Managed Agents webhooks endpoint and route the events into your own audit trail or dashboard.
Dreams: reorganize memory in the background
Announced in the May 6, 2026 release notes and extended on July 10 to support Claude Fable 5 and Sonnet 5, Dreams is a research preview that reads an existing memory store plus past session transcripts and produces a reorganized output memory store: duplicates merged, stale entries replaced, new insights surfaced. Endpoints are gated by the dreaming-2026-04-21 beta header; access is by request. Treat it as the API-side analogue of the app's monthly recap — a scheduled reflection pass that keeps the store from silting up.
App vs. API — one table
| Aspect | Claude.ai app | Managed Agents API |
|---|---|---|
| Storage unit | Categorized entry | Memory at a path |
| Grouping | Category (professional context, communication preferences, ongoing projects, …) | Path segments (/projects/acme/…) |
| Read/write | Live during chat | Agent tool calls during a session |
| Human-editable | Yes, per entry in Settings › Memory | Yes, via memory-store endpoints |
| Wipe | Per entry, or all | Per memory, per prefix, or delete store |
| Reflection | Monthly recap (Settings › Reflect) | Dreams (research preview) |
| Isolation | Incognito chat | Per-agent, per-session memory-store selection |
| Availability | Free, Pro, Max, Team, Enterprise (verify per plan) | Managed Agents (public beta) |
Where this lives in the rest of AILmanac:
- Memory & Context Editing — the client-side memory tool, a different beast: you implement the storage.
- Memory Across Chats — the beginner's overview of app memory.
- Agent Memory Architectures — the theory of episodic vs. semantic memory that makes sense of why categorized entries beat a rolling summary.
- Managed Agents — the harness that hosts the memory-store surface described here.
Privacy checklist before you turn any of this on
- Team and Enterprise: check whether admins can inspect entries. Treat the app memory as data classification 'internal', not 'confidential'.
- Never let Claude write API keys, tokens, or credentials into entries. If you paste one in a chat with memory on, delete both the message and any resulting entry.
- Use incognito for legal, medical, or financial one-offs. Nothing prevents the app from summarizing wording later if you don't.
- Audit periodically. Run the 'list every entry you have about me' prompt monthly and delete anything that no longer applies.
- On the API side, treat memory_stores as first-class data. Back them up, list-diff them before and after significant runs, and log memory_store.* webhook events to your audit trail.
Quick check
Check yourself
0/5Sources & further reading
- Claude Platform release notes — canonical source for the
agent-memory-2026-07-22beta header, SDK version bumps, memory_store webhooks (July 22), and Dreams (May 6, July 10). - Claude Apps release notes (Help Center) — categorized-entries rollout (July 10) and monthly recap (July 9).
- Use Claude's chat search and memory to build on previous context — official help article on the app memory model.
- Anthropic Newsroom — announcements for Claude for Teachers (July 14) and other July 2026 updates.
- Managed Agents documentation — memory store, webhooks, and Dreams reference.