Claude Agent Memory API Gets Its Own Beta Track: What the July 22 Header Migration Means for Operators

Anthropic's SDK v0.116.0 reveals that Memory Stores operations are moving from managed-agents-2026-04-01 to a dedicated agent-memory-2026-07-22 beta header — a breaking change every operator using cross-session memory must audit before July 22.

Published via Anthropic SDK

Archive item produced with AI assistance from the cited source and published without individual review. Editor of record: Joe Werner.

Abstract diagram of memory store API path branching from a single beta header to a dedicated memory header track

Anthropic's Python SDK v0.116.0 shipped on July 2 with a single feature commit that every operator using Claude Managed Agents Memory Stores needs to act on: the Memory Stores API is moving to a dedicated agent-memory-2026-07-22 beta header, replacing the existing managed-agents-2026-04-01 header that has governed all memory operations since April 2026.

This is not a deprecation warning. It is a header migration with a concrete deadline baked into the version string.

What happened

The SDK commit updates the anthropic-beta header injected by every method in client.beta.memory_stores.memories — create, retrieve, update, and list — from managed-agents-2026-04-01 to agent-memory-2026-07-22. The Memory Stores CRUD path (/v1/memory_stores/{memory_store_id}/memories) is the affected surface; session creation and session operations remain under managed-agents-2026-04-01.

Two API parameters are also removed in the same change: the order and order_by query fields on the list endpoint. Any integration that sends these parameters will receive a 400 after the header gate takes effect.

If you use the official Python SDK and you update to v0.116.0 or later, the SDK injects the new header automatically. No code changes required — just a package bump and a dependency pin check.

If you call the Memory Stores API directly via HTTP (curl, custom clients, non-Python SDKs), you need to manually update your anthropic-beta header before the July 22 activation date.

Why it matters for AI engineering teams

The memory stores capability was launched in April 2026 under the general managed-agents-2026-04-01 umbrella, which was always intended as a convenience flag that groups multiple in-beta agent API surfaces together. Graduating Memory Stores to its own header track signals that Anthropic is treating it as a stable, independently versioned API surface — a meaningful distinction for operators.

For teams running production agents, the practical implications are:

Cross-session persistence is now a first-class API surface. Moving Memory Stores to its own beta header is how Anthropic signals that this is not a throwaway experiment. Memory stores — workspace-scoped collections of text documents that mount into an agent session's sandbox — are a durable architectural commitment. Operators can build memory-dependent workflows with more confidence in API stability.

Header management becomes more granular. You can now opt into memory operations independently of the general managed-agents surface. This matters for teams that want to scope beta exposure: if you are using Memory Stores but not yet consuming every managed-agents feature, you can restrict your beta surface by sending agent-memory-2026-07-22 without necessarily depending on managed-agents-2026-04-01 for memory calls.

The list endpoint simplification removes ordering flexibility. The order and order_by parameters being dropped from memory listing means you get memories back in the store's natural order. If your agent tooling or memory management dashboard sorted memories by a custom field, you need to sort client-side after July 22.

The router/operator angle

For teams routing agent traffic through a gateway or building multi-agent systems, the Memory Stores migration exposes a pattern that will repeat: Anthropic beta headers are API versioning in disguise. Each anthropic-beta flag is effectively a versioned surface contract, and when a surface graduates to its own header, the old shared flag stops being authoritative for that path.

Operators should build their SDK wrappers and gateway configurations to handle header routing explicitly rather than relying on SDK defaults. The correct pattern:

  • Session creation and operations: managed-agents-2026-04-01
  • Memory Stores CRUD (after July 22): agent-memory-2026-07-22
  • Dreams research preview: dreaming-2026-04-21 (additional, alongside managed-agents header)

If your gateway proxies Claude API calls and injects beta headers at the request level rather than letting the SDK manage them, this is the moment to audit your header injection logic. A proxy that injects managed-agents-2026-04-01 on all /v1/memory_stores/ paths will likely receive 4xx errors once the new header requirement activates.

For teams using the Python or TypeScript SDKs with automatic header management, the only operational risk is pinning an older SDK version. If you're on anthropic<0.116.0 and the July 22 cutover happens, memory store operations will send the old header and fail. Pin to >=0.116.0 explicitly.

What operators should audit before July 22

A targeted three-step audit before the deadline:

  1. Search your codebase for managed-agents-2026-04-01 in request headers, config files, gateway middleware, and API wrapper code. Any reference that governs /v1/memory_stores/ calls needs to update to agent-memory-2026-07-22.

  2. Check your SDK version pins. If you use the official Python SDK, bump to >=0.116.0 and verify that your anthropic-beta parameter doesn't manually override the header for memory calls (the SDK merge logic would otherwise combine old and new flags).

  3. Audit the list endpoint consumers. Remove any order or order_by parameters from calls to GET /v1/memory_stores/{id}/memories — the parameters are being dropped server-side, and sending them post-cutover will cause parameter validation errors.

Teams that build stateful multi-agent workflows on top of Anthropic Memory Stores — where each agent session inherits pre-populated context, user preferences, or project conventions — have a 19-day runway to test and validate. Don't wait until July 22.

What TheRouter users should watch

If you route Anthropic-compatible requests through TheRouter and your agents use the Managed Agents Memory Stores API, check whether your routing layer injects or strips custom anthropic-beta headers. TheRouter passes through provider-specific headers for direct Anthropic endpoints; if your integration relies on that pass-through, validate that the updated header reaches the Anthropic API correctly after a dependency bump. Review the TheRouter docs for Anthropic provider configuration and header forwarding behavior.

Help & contact