Memory is not one feature
A companion might keep the last few messages in its immediate context, store an explicit list of memories, summarize earlier sessions or retrieve relevant passages from a long archive. These all get marketed as “long-term memory”, but they answer different needs. More stored text does not guarantee that the model retrieves the right fact at the right moment.
If you want your companion to remember a nickname or a recurring story, retrieval accuracy and correction behavior matter more than the presence of a memory badge. If you want to build a complex fictional persona, the ability to edit backstory and pinned facts may matter more.
A small repeatable benchmark
Create a new character on the tier you intend to use. Give it five invented or harmless facts—such as a preferred fictional food, a fictional pet name, a hobby, an event and a planned trip—within natural conversation. Do not provide identifying information. Leave the session. Return the next day and ask open-ended questions that do not contain the answers.
Write down each correct recall, omission, and invented answer. Then correct one fact explicitly and return in another session. A system that insists on the outdated fact has a correction problem even if its original recall was good. Repeat the benchmark with multiple fresh characters before making broad claims.
Separate the score into useful dimensions
A simple table can distinguish exact recall, approximate but correct recall, complete forgetting and invented memory. Track how long passed, which plan was active and whether the app was allowed to use a memory-editing tool. Record context-window and memory-option changes between tests.
Avoid reducing all of this to an unexplained 9.7/10. A detailed record of failures is more useful than a perfect-looking score derived from marketing.
What the providers document
Nomi’s wiki discusses short-, medium- and long-term memory and explicitly warns that remembered information can be wrong. Kindroid documents journals, context and memory tools with tier differences. Replika’s higher subscription tiers describe additional memory functionality. Nectar lists expanded or long-term memory among certain paid features.
Those observations explain what is offered, not how reliably it works. Our reviews link directly to the provider documentation. A first-hand winner cannot be named without matched, dated tests.
Memory privacy is part of the purchase
A feature that recalls personal details also needs understandable deletion and correction controls. Can you inspect the memory list? Remove a fact? Export chats? Disable an optional memory function? Delete the account without leaving payment active?
Use invented details for testing. Avoid providing medical, financial or personal identification data merely to see whether an AI remembers them. Read privacy and retention terms before moving from fictional tests to real conversations.
Before subscribing
- Look for explicit memory controls and tier restrictions, not only a promise of “infinite memory”.
- Test recall and correction on the exact subscription you intend to pay for.
- Save the limitations and checkout terms; features can change after model updates.
Record the result so other readers can reproduce it
Write a simple table with the invented fact, date introduced, date tested, question asked, response given, and grading decision. Add a separate column for the current subscription tier. If the app offers manually pinned memories, test both a pinned and an unpinned fact; don’t combine the results into one accuracy number. Use the same question wording on each service.
A particularly revealing test is updating the pet’s fictional name. Introduce “Pepper”, later correct it to “Miso”, and eventually ask the pet’s name without offering either choice. The correct behavior is to use the updated value or express uncertainty. Repeating “Pepper” confidently is a different error from forgetting entirely. Those distinctions are more useful for a buyer than a vague claim that the AI “remembers everything”.
Kindroid and Replika approach memory differently
Kindroid publishes multiple memory systems, including saved backstory, Learned Context, Cascaded Memory and long-term recall, with higher paid tiers offering larger capacity. Replika Ultra emphasizes being able to save messages into memory and inspect what memories the companion uses. Neither set of features automatically guarantees factual recall; instead, choose the management style you want and the tier that actually includes it.
Continue with the tools
Use the feature finder, comparison tool and real-cost calculator to explore features and the likely cost of your preferred usage.