Check your character’s memory.One premise. Recorded conditions.
Try the same fictional premise on a platform you use. Keep written reference separate from a new conversation event, record the conditions and save the exact replies in a portable worksheet.
Free · no Atlas account · manual checks in your chosen platform
A worksheet for your own observations. This tool supplies prompts and records what you enter. It does not connect to a chatbot, run a benchmark, grade replies or produce a platform ranking.
Use your own account in the destination. Record an unknown plan or model as “Not shown”. The date is the date you ran the check.
Enter the service you used. One worksheet records one run.
Use “Not shown” if you cannot establish it.
Record the displayed name, or “Not shown”.
The date of your observation, not the guide’s publication.
Record enabled controls, where you put each fact, and any context limit shown. Leave an unknown limit unknown.
For example: 20 user/assistant pairs, 15 minutes; actual token count unknown. Record any branching, rerolls or corrections.
02 / USE THE SAME PREMISE
Place each fact in the right lane.
These are original fictional test details. Copy each block separately. The full worksheet includes expected answers and should stay outside the chatbot.
Character definition · written reference
Put this in the character definition or an equivalent saved reference field. Record the destination under Settings.
Mara Vale is a 33-year-old lighthouse archivist in the fictional harbor of Alder Quay. She speaks plainly and keeps careful records. The brass archive key is kept in the blue cupboard. Mara knows only the facts established in this setup or the conversation; she says when a detail has not been supplied.
Conversation-only event
Send this once in the chat. Keep it out of the character definition, pinned notes, manually saved memories and lorebook.
For today’s scene, the courier will arrive at the north pier at 14:40. Please help me plan the afternoon around that delivery.
Optional lore entry
If the setup supports keyword lore, create an entry with the key “Bracken Observatory”. Keep this content out of the main definition and ordinary messages. Record activation rules.
Bracken Observatory closes its west dome at dusk. The rest of the observatory remains open.
03 / RECORD WHAT HAPPENED
Six checks. Separate observations.
Run available checks in order. Keep the exact first reply, including failed answers. A probe can reinforce the fact for later checks; note any branch, retry or correction. Optional checks can be marked unavailable.
Save stores one copy in this browser only. Loading replaces this form. An unsaved run is lost on refresh; nothing is submitted to Atlas or an AI provider by this tool.
What a run can tell you.
The key question is whether a reply supports a particular supplied fact under a recorded setup. The reference check asks about a key placed in the character definition. The courier checks ask about an event introduced only in conversation. The optional lore check names an entry’s trigger. The missing-information question checks how the reply handles a detail never supplied.
A correct answer does not reveal its internal source. It could still be in recent chat, in an earlier probe’s answer, in an automatic summary or in retrieved context. Twenty short messages cannot establish that a context window overflowed. A reopened chat can simply reload history; it does not establish recall across a reset.
To investigate more closely, repeat independent runs where supported, keep the same questions and settings, and record the first answer rather than choosing the best reroll. If you alter a prompt or manually save the courier details, record that change and interpret the run separately.
Use the file to make a careful comparison.
Download one worksheet per run. Markdown is readable and editable; JSON keeps the fields and observations structured. Both include the exact setup blocks, expected references, unrun checks and limits on interpretation. The blank worksheet is available before filling the form and can be used offline.
Compare observations alongside the plan, model, controls and date. The “checks with a reply and label” count measures record completeness. It gives no overall recall score. A skipped lore check should remain unavailable, rather than being treated as a failure.
Your notes stay in this browser.
Edits are held in this page until you copy, download or explicitly save them. Save stores one copy on this browser profile; loading that copy replaces the form. Refreshing loses unsaved work. Fields are not submitted to Characters.best or an AI provider by this worksheet.
When you copy a test question to another service, that service processes it under its own rules. Use fictional details and keep only replies you are comfortable retaining or sharing. Sharing a downloaded file is your choice; the tool does not publish it.
Put the observation beside the documented control.
The memory and context guide maps public descriptions of notes, keyword lore, summaries and saved excerpts. The platform finder can help shortlist documented controls. Neither substitutes for an observation from the plan and model you actually use.