A practical way to compare companion memory claims is to ask who can see and correct them

Use a control-based comparison

A useful comparison starts with four questions: what is always present, what is retrieved only when relevant, what can the user inspect, and what can the user change or remove? Kindroid’s current documentation answers those questions with persistent setup, cascaded medium-term continuity, and retrievable long-term or journal material. It says recalled retrievable items can sometimes be seen on a message, while some system details remain invisible. Replika’s help center describes a visible Memory tab where users can add, edit, or delete individual facts, alongside a deeper pattern-based layer.

This method is more informative than ranking products by a single advertised memory number. A creator may value a durable backstory; a roleplayer may need long-running continuity; a privacy-conscious user may prioritize inspection and removal controls. These are reader priorities, not quality scores. Documentation language also changes, so a comparison should record retrieval date and avoid translating a plan name into a capability the company has not documented.

Limits of the method

Evidence: both help centers document layered memory and at least some user controls. Inference: inspectability and editability are practical proxies for agency, not proofs of complete control over all stored or inferred data. They also do not tell us how accurately a product retrieves context or how well it fits a given person. Before paying or sharing sensitive material, readers should review the current in-product controls and privacy documentation directly; this note does not recommend a service or provide a test score.

Evidence limit: This is a framework derived from documentation, not a feature-complete market comparison or an endorsement.