After a long session, Claude forgets an exact order number from early on. What protects it?
Why D is the answer
Keeping key facts in a pinned block that is resent with every turn
Summaries lose exact values. Keep facts that must survive in a structured block sent on each turn rather than trusting history.
In detail
- Summaries of long conversations can lose exact values, such as order numbers.
- Keep facts that must survive in a structured block, and send it on every turn.
- They then stay in context no matter what happens to the older history.
- Update that block when a fact changes, and keep it short.
Use case
An illustrative scenario, not a real incident.
A long support session summarizes its earlier turns, and order 48213 becomes "the customer's order". A short block of facts, sent on each turn, keeps the exact number available.
See it in different ways
stateDiagram-v2 [*] --> CheckInput CheckInput --> Rejected400: input too long CheckInput --> Generating: input fits Generating --> MaxTokens: reply limit reached Generating --> WindowFull: model_context_window_exceeded Generating --> Finished: end_turn
flowchart LR S0["Key fact found"] --> S1["Added to pinned block"] --> S2["Sent each turn"] --> S3["History can be summarized"]
Key point: Keep exact values in a block that is sent each turn.
Why the other options are wrong
- A. Asking Claude to try harder to remember the early part of the conversation
- If the value has left the context, asking harder cannot bring it back.
- B. Raising the temperature so that the model explores more of its earlier turns
- Temperature changes variety in replies, not what is in context.
- C. Starting a new session once the conversation has reached a certain length
- A new session loses even more unless the fact is carried over.