Claude Managed Agents Memory API Changed Today: Fix These 3 Things Now

Anthropic changed the behavior of every Claude Managed Agents memory call in production on July 22, 2026, with the new `agent-memory-2026-07-22` beta header now live and the old `managed-agents-2026-04-01` header adopting the same new list semantics. Developers passing custom `depth` values or using `path_prefix` without a trailing slash will hit 400 errors or silent regressions, requiring immediate fixes to three key parameters: `depth` now only accepts 0, 1, or omission; `path_prefix` requires a trailing slash and matches whole path segments; and `order_by` and `order` parameters are silently ignored.

Anthropic changed the behavior of every Claude Managed Agents memory call in production today. The agent-memory-2026-07-22 beta header is now live — and the critical part is that you do not need to switch headers to be affected. As of July 22, the old managed-agents-2026-04-01 header adopts the same new list semantics. Developers passing custom depth values or using path prefix without a trailing slash will hit 400 errors or silent regressions starting now. Here is what changed and how to fix it. Why Claude Agent Memory Matters Claude Managed Agents Memory launched in public beta on April 23, 2026 https://claude.com/blog/claude-managed-agents-memory . A memory store is a workspace-scoped collection of text documents mounted inside a session container at /mnt/memory/<store-name / . The agent reads, writes, and updates them using the same file tools it already uses for code — no special memory API inside the agent, just the filesystem. The results speak for themselves. Rakuten’s agents report 97% fewer first-pass errors, 27% lower costs, and 34% lower latency — because agents that remember every past mistake stop repeating them. Netflix, Wisedocs, and Ando are running production workflows on it. The April launch was three months ago, and adoption is ahead of the announcement pace. Today’s API change reflects that: the memory subsystem is being tightened for GA readiness. Three Things That Break Today 1. The depth Parameter Is Now Strict Previously, you could pass any integer as depth on a memory list call and the API would handle it often silently clamping or ignoring the value . Starting today, the only accepted values are 0 , 1 , or omitting the parameter entirely. Any other value returns a 400 error. If your code does this, it breaks today: Breaks with 400 starting July 22 memories = client.beta.memory stores.memories.list store id=store id, depth=2 Correct — use 0, 1, or omit entirely memories = client.beta.memory stores.memories.list store id=store id, depth=1 2. path prefix Now Matches Whole Segments, Not Substrings This one is more dangerous because it will not throw an error — it will silently return wrong results. The path prefix parameter now requires a trailing slash and matches whole path segments, not substrings. path prefix="user" previously matched user-data/ , user-preferences/ , and user/ . After today it matches nothing — it lacks the trailing slash. path prefix="user/" now matches only the user/ directory and its contents. Silent regression — may return empty or wrong results starting today memories = client.beta.memory stores.memories.list store id=store id, path prefix="user" Correct — trailing slash required memories = client.beta.memory stores.memories.list store id=store id, path prefix="user/" Audit every path prefix call in your codebase. If it does not end with / , it is now broken. 3. Result Order Is No Longer API-Controlled The order by and order parameters are now silently ignored. Results return in a stable server-defined order. If your code relies on the API returning memories in a specific order — by creation time, alphabetically, or otherwise — review that logic now. Who Is Actually Affected If you are using a current SDK, you are partially protected. The Python SDK 0.116.0+ https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.116.0 , TypeScript SDK 0.110.0+ , Go 1.56.0+ , Java 2.48.0+ , Ruby 1.55.0+ , PHP 0.36.0+ , C 12.35.0+ , and CLI 1.16.0+ all send the correct header automatically. But they do not protect you from the depth and path prefix behavior changes — those apply regardless of SDK version or header. If you call the API directly or set beta headers manually, replace managed-agents-2026-04-01 with agent-memory-2026-07-22 on memory store calls. Sending both headers on the same request returns a 400. Also: page cursors issued under the old header are not valid under the new one — restart pagination from page 1 when adopting the new header. The Three-Step Fix Step 1: Upgrade your SDK. Python pip install "anthropic =0.116.0" Node.js npm install @anthropic-ai/sdk@latest Step 2: Search your codebase for path prefix and depth arguments on any memory store list call. Add trailing slashes to every path prefix value. Replace any depth value above 1 with 1 , or remove the parameter entirely. Step 3: If you set beta headers manually, swap managed-agents-2026-04-01 with agent-memory-2026-07-22 on memory store calls only. Non-memory-store Managed Agents calls still use managed-agents-2026-04-01 . The full migration reference is in the Claude Platform memory docs https://platform.claude.com/docs/en/managed-agents/memory . The developer platform release notes https://platform.claude.com/docs/en/release-notes/overview have the complete changelog. The API tightening is a clear signal that memory is heading toward GA — APIs do not get stricter validation as a courtesy. Start treating your memory store calls with the same rigor you apply to any production endpoint.