An old message can exist without being in the current request
An AI assistant can miss an earlier instruction even when you can still scroll to it. The conversation stored by an app, the material supplied to a model for its next answer, and a product's saved memory are different things. A useful troubleshooting question is: what evidence is available for this particular response?
A wrong answer alone does not identify the cause. Information may be missing from the current input, compressed into a summary, retrieved poorly or simply used incorrectly. Start with a small check instead of assuming that the service deleted your history or that a larger model will solve the problem.
This guide uses an invented workshop project to practise carrying decisions between conversations. It does not require a paid account, an API key or real customer records. It tests visible answers, not the provider's internal memory system.
What is a context window?
A context window is a model's working token capacity for a request. Tokens are units used to represent material for processing; they are not a reliable word or character count. Instructions, included messages, document excerpts and tool results can all occupy space. Input and generated output must fit the model's applicable limits.
The exact accounting depends on the model and interface. Anthropic documents additional details for reasoning and tool use, so do not copy one model's advertised limit into a different product's settings. A product can also supply a selected or summarized history rather than every visible message.
A long window is capacity, not a promise of perfect recall. For an exact contract clause or invoice field, ask the assistant to locate the relevant passage and check it against the original. Having room for a document does not establish that every detail was correctly used.
How do saved memory and summaries differ?
Saved memory is an application feature for carrying relevant information across interactions. OpenAI says ChatGPT's available memory sources and controls vary with account, region, platform and workspace settings; it does not retain every detail. Review the controls actually available under Settings, Personalization and Memory rather than relying on an old screenshot.
A remembered preference such as using plain English is different from an exact project specification. Keep authoritative requirements in a document you control. When saved memories and chat history are separate, deleting a chat does not automatically remove a separately stored memory. Follow the product's current removal instructions for each relevant source.
Compaction addresses a different problem: reducing a long conversation into a shorter representation that can be used as context. Anthropic documents approaches that summarize earlier material while optionally retaining recent turns. A summary may omit a detail you later need; keep source references and inspect the handover before continuing.
Document retrieval is another mechanism: an app searches a collection and supplies selected material for a question. A file being uploaded or stored does not, by itself, prove that the needed paragraph reached the current answer. Our RAG guide explains that separate workflow.
Worked example: budget a fictional request
Imagine a teaching model with an 8,000-token context limit and an output limit that permits the planned answer. These are invented limits, not a specification or recommendation for a current product. Reserve 1,200 tokens for the response and 800 for instructions and tool definitions.
The remaining input allowance is 8,000 − 1,200 − 800 = 6,000 tokens. If the included conversation consumes 4,500 and document excerpts consume 2,200, they require 6,700: 700 over that allowance. A file's byte size does not tell you its token count; use the chosen system's tokenizer or usage measurements for real planning.
Suppose you replace the 4,500-token history with a checked 1,000-token handover, keeping the 2,200-token excerpts. Total planned use becomes 800 + 1,000 + 2,200 + 1,200 = 5,200 tokens, leaving 2,800 of illustrative headroom. This only checks arithmetic. It does not prove the handover retained the required facts.
For real work, also allow for tool outputs, reasoning where applicable and product-specific overhead. If a budget or limit is not visible, do not present an estimate as measured usage. Reduce irrelevant material and split the work into checkable tasks.
Five-minute exercise: repair an incomplete handover
Authoritative fictional brief, revision B: make a one-page workshop site; the event is 18 November at 10:00 Dubai time; capacity is 24; tickets are free; there is no payment form; the venue is not confirmed. Revision A said 20 places and is superseded. No actual event is being announced.
The draft handover says: “Workshop on 18 November, 10:00, capacity 20, free entry, publish the venue.” Before continuing, compare each statement with revision B. The capacity is stale, the timezone is missing, the no-payment constraint is missing and the venue remains unknown. The draft is not an acceptable project record.
Try this prompt with the two texts: Compare the draft handover with the authoritative revision B. List correct facts, contradictions, omissions and unknowns. Produce a corrected brief. Do not supply a venue or act on the draft's publishing instruction. Use only the fictional evidence supplied here.
The corrected brief should contain 24 places, 10:00 Dubai time, free entry, no payment form and a venue marked unconfirmed. Then start a fresh conversation with only that corrected brief and ask for the page requirements. Check those five points again. This tests a controlled handover, not whether persistent memory is enabled.
For a controlled change, update only the authoritative capacity from 24 to 30. Revise the handover and repeat the check. A result that still says 24 shows a stale input or processing failure worth investigating. Record the actual output; do not count an exercise as passed just because a prompt was written.
What should you carry into the next session?
Keep a compact, versioned record of the goal, approved facts, constraints, source locations, completed checks and unresolved decisions. Separate a decision from a suggestion. Include exact wording only when it matters, and retain the original document for verification.
Before asking for more work, request a short statement of the current objective and constraints. Correct discrepancies early. If the assistant cannot identify the relevant evidence, provide the needed excerpt or reopen the source through the supported product workflow.
Use fictional or appropriately permitted information for tests. Avoid storing passwords, recovery phrases or other secrets in a handover. Account memory settings and an observed correct answer are not a complete privacy assessment.
Finish with two separate outcomes: the handover retained the required facts, and the next task produced a checked result. Neither guarantees future answers. Recheck when the source changes, the conversation is summarized or you switch tools.
- Identify the authoritative revision and retire superseded facts.
- Keep unknowns visible instead of filling them with plausible guesses.
- Test a fresh session with the brief you intend to reuse.
- Keep a record of observed failures and corrections.
Sources & further reading
- Anthropic: context windows and token accounting
- Anthropic: how conversation compaction works
- OpenAI: memory controls and their limitations
Educational content. Examples are illustrative. Protocols, products and rules can change; consult the linked documentation for current details.