AI Girlfriend Apps Rarely Disclose Memory Persistence

As of September 16, 2026, across the 36 platforms we track in AI Girlfriends, only 36% of persistence scores rest on a platform's own documented policy, while 56% had to be inferred from feature lists that never explain how conversation memory is stored, reset, or controlled. The average score across all 36 platforms is 4 out of a possible range that runs from 1 to 8.5. Only 8% of scores come from direct, evidence-based observation of the product in use. The gap between platforms that clearly state what happens to a user's chat history and those that merely imply it through marketing language is the defining feature of this category right now, and it means most users are guessing at how persistent their AI companion's memory actually is.

Thirteen platforms explain persistence in their own documents

The strongest scores in this category belong to platforms that publish specific, checkable details about memory architecture. Erogen leads at 8.5, with a justification describing tiered context memory of 4K, 8K, and 16K tokens, core memory slots ranging from 0 to 24, and chat history retention windows from one year on the Free tier to unlimited on Silver and Platinum plans. Girlfriendly AI scores 8.0, documenting model memory up to 128K tokens on its Premium and Deluxe tiers and up to 1M tokens on Ultimate, paired with explicit chat profile caps of 50, 100, and 200. Several platforms cluster at 7.5: Candy AI states plainly that it offers 'text-based chat with memory retention'; DarLink AI ties memory tiers to plan level, from basic memory through 'living memory' on its Ultimate tier; My Dream Companion advertises 'Long-Term Memory' with an 'Extended Chat Memory' paid upgrade; and Pornify documents three named tiers — Basic, Enhanced, and Living memory. These seven platforms, plus Secret Desires AI at 7.0, account for most of the 36% declared share, and they demonstrate that clear documentation is achievable when a platform chooses to provide it.

Declared claims without user controls still score in the middle

A second cluster of declared scores shows that advertising memory is not the same as making it manageable. AIpornX scores 6.0: its feature list promises 'private conversation history' and testimonials claim the AI remembers past chats, but no controls exist for viewing, exporting, or deleting that history. Fanfinity, also at 6.0, documents 'Memory and Personalization' and states the companion 'remembers past conversations, shared moments, and your preferences,' yet offers no way to reset or inspect that stored data. Infatuated.AI and Moescape share the same 6.0 score for comparable reasons — the latter's privacy policy references processing data to 'maintain conversation context and memories' and mentions 'lore books for roleplay context,' which counts as documentation but stops short of giving users agency over it. Soulkyn rounds out this tier, also at 6.0, advertising memory and adaptation across sessions without reset options. JOI, by contrast, scores only 3.0 despite being declared: its privacy policy notes that chat facts and profile information are stored to personalize conversations, but the absence of any user-facing deletion or reset mechanism pulls the score down even though the underlying practice is disclosed. This spread shows that within the declared category itself, disclosure of a feature's existence and disclosure of user control over it are scored as separate things.

Observed testing exposes gaps between marketing and delivery

Only three platforms in this category carry an observed score, meaning our reviewers verified behavior directly rather than relying on stated policy. TryNectar AI scores 3.5 after independent testing found that memory 'tops out at 35 messages on Pro tier,' even though long-term memory with 64K context is advertised only on the Ultimate tier — a documented mismatch between marketed capacity and delivered capacity. Texthub scores 2.0 on the basis that its chat interface shows a 'Restart chat' control and an online status indicator per character, but no documentation anywhere describes how history is stored or whether it persists at all. SpicyChat scores lowest among observed platforms at 1.0, where reviewers found no documentation on chat history, memory controls, or session persistence, and the platform's own features list returned empty during assessment. These three cases matter disproportionately because they are the only points in the dataset where a claim was checked against actual behavior rather than taken at face value, and in two of the three cases the check revealed a shortfall.

What to verify before paying for long-term memory

The 56% of platforms scored as inferred — including Angel AI at 4.0, ourdream.ai and Privee AI at 3.5, and a large group of platforms scored 2.0 to 3.0 such as Kupid AI, Luvr AI, Promptchan AI, SoulFun, DreamGF.ai, Fantasy.ai, and Romantic AI — advertise ongoing companionship or persona features without ever specifying how conversation memory is stored, exported, or reset. For a user paying specifically for a companion that remembers past interactions, this is the category where expectations are least likely to match reality. Before subscribing to any platform on the basis of a 'memory' or 'long-term companion' claim, it is worth checking:

• Whether the platform names a specific memory unit (message count, token limit, or time window) rather than vague language like 'remembers you'
• Whether memory capacity is tied to a paid tier, and if so, what the free or entry tier actually retains
• Whether there is a visible control to view, export, or delete stored conversation history
• Whether any independent testing or user reports confirm the advertised memory limit in practice
• Whether the privacy policy, not just the marketing page, addresses data retention

Given that 20 of 36 platforms in this dataset carry an inferred score, the honest baseline assumption for most of this category is that persistence behavior is undocumented until a specific platform proves otherwise.