Claude Preserved Thinking: Audit Conversation Replay
Check conversation storage and replay after Anthropic’s preserved-thinking changes. Separate display summaries from API history and find integration assumptions before rollout.
Published · 4 min read
At a glance
- Display formatting and API replay need separate representations.
- Check account, organization and model boundaries explicitly.
- Use documented context management rather than editing protected history.
The integration issue behind the news
Audit how your application stores and replays conversation history. Anthropic’s October 12 report says preserved thinking, introduced with Fable 5.1 and Opus 5.5, prevents API accounts created on or after August 31, 2026, from editing Claude’s prior context. It also describes organization restrictions for Sonnet 5.5 reasoning transcripts. Read the report and follow its links to current implementation documentation.
The practical issue for legitimate applications is compatibility. A chat interface may shorten or reformat messages for display, while an API conversation requires a provider-supported representation. Treating the rendered transcript as a complete replay log can break continuation even when the visible conversation looks correct.
Inventory every transformation of history
Trace one synthetic conversation from provider response to storage, UI rendering, background job and resumed request. List any step that drops blocks, edits fields, merges messages, changes roles or transfers context to another account. Include observability middleware: a log sanitizer may be safe for a support view yet unsuitable as the source for an API replay.
Keep credentials out of this inventory. Record account creation cohort and organization as nonsecret identifiers, and inspect the exact model used. A test under one older account is not proof that a newly provisioned organization will behave identically.
Separate presentation from protocol state
Use a dedicated representation for the user-visible transcript. It can contain concise summaries and redacted display text. Keep the supported continuation state separately, under your normal security and retention controls. Do not assume that every protected block is meant to be readable, transferable or editable.
When privacy policy requires deletion or a fresh boundary, use the provider’s documented new-session or context-management mechanism. Do not try to reconstruct, forge or bypass protected state. A new conversation with an explicit human-readable handoff can be a valid application design, but it is a new context and should be labeled accordingly.
Build a small compatibility matrix
Run only authorized, synthetic conversations. The following matrix is a proposed integration test plan, not a set of observed provider results.
| Case | What to check |
|---|---|
| Continue within the same supported session | Stored state resumes without app-side mutation |
| Resume after a normal app restart | Persistence preserves required fields and ordering |
| Display a redacted support view | UI changes do not overwrite continuation state |
| Change account, organization or model | Application follows documented support boundaries |
| Start a new context from a summary | App labels the handoff and accepts lost detail |
Record exact errors and request IDs using your approved logging policy. Avoid pasting protected content or private user conversations into a bug report.
Ask AI to inspect the application, not protected reasoning
Use a coding assistant to review your storage adapter with synthetic fixtures:
Review this redacted conversation persistence and replay code:
{{adapter_code_and_synthetic_fixtures}}
Identify transformations between API response, storage and UI.
Return file | transformation | protocol assumption | test needed.
Do not extract hidden reasoning or bypass provider protections.
Mark undocumented behavior UNKNOWN and link it to a question
for the current official API documentation.
The useful output is a concrete adapter finding, such as a display formatter incorrectly overwriting stored history. It is not a claim that all transcript editing is safe or that all models share one replay format.
Acceptance check before rollout
Confirm that your supported continuation path passes after a restart, UI redaction cannot mutate protocol state, and cross-boundary cases fail clearly or use a documented handoff. Keep a versioned synthetic fixture with the client and model details. If the documentation leaves a boundary unclear, resolve it with the provider before depending on it.
For a portable human-readable task summary, use the agent handoff guide. Keep that convenience separate from claims about lossless API replay. Your next step is tracing one conversation through your own code and identifying the source of truth for continuation.
Sources and editorial review
Reviewed on 22 October 2026. Primary-source reporting and editorial analysis; the exercises and example results are illustrative, not hands-on performance measurements. Product access, regional rollout and commercial terms can change.
01 Choose a tool: compare the fit. 02 Use a prompt: start with context. 03 Check the result: keep your judgment.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkReview evidence
Documentation-based review · 2026-10-22
Primary sources checked on 22 October 2026. News-led editorial selection, with access limitations recorded in the batch research notes. No measured exact-query popularity, comparative testing or traffic outcomes are claimed.
