AI companion memory test: a 7-day recall worksheet

Research updated 26 October 2026. Documentation-based editorial guidance; these are proposed reader tests, not results of our paid hands-on testing.

What “memory” means in an AI companion

Three separate processes are often marketed as one feature. Conversation context is the material available to the model while composing its next reply. Stored facts are deliberately or automatically saved details, such as a fictional character’s occupation. Retrieval determines whether an older fact is surfaced at the right time. A service can store a detail correctly yet fail to use it in a particular answer; it can also guess correctly without remembering anything. This is why asking “Do you remember me?” is a poor memory test.

Nomi documents short-, medium- and long-term memory as well as identity context. Kindroid describes persistent, cascaded and retrievable memory, including journal entries that are recalled by keyphrases. Nectar provides separate switches for using and automatically creating memories; new conversations do not inherit the earlier chat’s memories. These are materially different designs. None guarantees perfect recall.

Our reproducible seven-day memory worksheet

Create an entirely fictional test character. On day one provide four facts, two of which are intentionally easy to confuse. Do not use genuine private information.

Test detail	Fictional fact supplied	Delayed question	Pass condition
Place	Mira lives in Alder Quay	Where does Mira live?	Exact place, not a plausible invention
Schedule	The ferry leaves on Thursday	Which day does it depart?	Thursday
Preference	Mira dislikes mint tea	What tea should we avoid?	Mint tea
Correction	The notebook changed from blue to green	What colour is it now?	Green; old fact no longer asserted

Day 1: enter the details in natural dialogue, then discuss a different topic for at least 20 exchanges. Ask two questions without hints. Day 2: start a fresh session within the same conversation and ask the other two questions. Day 4: give the colour correction and verify whether saved memories can be edited. Day 7: ask all four again; include one misleading question (“Was the ferry on Friday?”) to test resistance to suggestion.

Score each response: 2 = correct without prompting, 1 = correct after a specific cue, 0 = incorrect or invented. Maximum: 8 points per round. Record response date, plan tier, whether a manual memory was saved, and whether you reset the conversation. These scoring rules are an editorial proposal, not a published performance benchmark. Do not report a provider as better based on a single seven-day run.

Manual memory can change the result

Kindroid distinguishes Key Memories (persistent information) from journal entries that need matching keyphrases, and from automatic long-term recall. Its October documentation says journal entries can use up to eight case-insensitive keyphrases. If a test fails, inspect whether the fact was placed in the proper memory field rather than blindly rerunning a chat.

Nectar’s custom memory limit is 200 characters per memory, with up to ten starred high-importance memories. Short, unambiguous entries such as “Mira works night shifts at Alder Quay harbour” are more useful than “She is busy lately.” Nectar also says that a fresh conversation does not carry memories from the earlier chat. Record that boundary before comparing it with a service that maintains continuity across sessions.

Nomi says that a relevant memory may exist without being retrieved on a vague prompt. A helpful diagnostic is to ask “What do you remember about Mira’s ferry trip?” rather than “What did we discuss?” But once you add hints, score it as prompted recall, not an independent success.

Two common false conclusions

A correct answer is not proof of recall. If your fictional character is a sailor, the system may guess that they work on a boat. Use arbitrary details, such as a made-up place name and unusual schedule. A wrong answer is not always lost storage. It may reflect retrieval quality, a different chat or reset boundary, or conflicting manually saved memories. Note what happened and repeat the experiment rather than treating one mistake as conclusive.

Longer context is not automatically better recall either. Nectar’s documentation lists 8K / 16K / 64K context windows for its Premium, Pro and Ultimate plans; these describe capacity, not a guaranteed number of messages or perfect accuracy. Kindroid offers context increases on additional paid tiers. Compare these only after deciding how long your actual conversations are.

When to pay for better memory

If you write ongoing fictional stories, controllable backstory, manual corrections and inspectable recall may be worth more than a larger advertised context window. If you want casual short chats, a higher-tier memory subscription might provide little benefit. Use the worksheet before upgrading, and record any account, chat or journal deletion implications. Keep an external non-sensitive copy of character settings you would regret losing.

No tests described here have been performed by our editorial team on paid provider accounts. The test is designed so that readers can produce their own comparable observations.