Stay up to date with AI tools, model releases, and developer workflows that matter.
It has been revealed that weekly. Free. One click to leave.
7 Day Free Trial. Cancel Anytime.
AI companions are the most demanding consumer use of large language models right now. A coding assistant can forget your last message and still be useful. A companion that forgets your name, your running story, or the strategy you made last Tuesday has failed at its one job. On top of that, owners now expect voice calls that feel like a phone conversation, not a walkie-talkie exchange with a five-second pause.
Recent findings suggest that three engineering problems decide whether a companion app works: how much of the conversation fits in the model's context window, what the app remembers once the window fills up, and how fast the voice pipeline answers. Here is how consumer companions handle each one, with real tier and pricing examples, and what developers can take from their designs for their own products.
According to recent reports, most LLM products are judged on single answers. Companions are judged on continuity. That creates pressure in three places at once:.
Every consumer companion is a different set of trade-offs across those three constraints, and the trade-offs are increasingly how the apps are priced.
Sources indicate that the context window is the total number of tokens the model can see when it writes the next reply. In a companion app, that budget is spent before the user indicates anything:.
When the chat log outgrows what is left, the oldest messages fall out. The user experiences this as the companion "forgetting" mid-conversation. This is why context size is now a headline spec in the category, and a paid feature. SpicyChat AI, for example, gives free accounts a 4K context window, then scales memory to 8K on its $19.95/mo tier and 32K on its $29.95/mo tier. Janitor AI runs about 9K of context on its own JanitorLLM. Independent uncensored AI chat benchmarks record these context sizes alongside refusal behavior and pricing, because context length predicts how long a roleplay stays coherent better than any marketing claim.
Some apps sell memory in messages rather than tokens. Fantasy GF's pricing page lists a 60-message base "smart memory" and 5x that, 300 messages, on its Ultra blueprint. Others sell it as a memory multiplier on higher subscription tiers. The pattern is consistent across the category: memory depth is the upgrade lever, not the model itself.
Industry observers note that no consumer app keeps an entire relationship in the prompt. Persistent memory is a separate system that writes to storage and reads back into the context window on each turn. In practice it has three layers:.
Industry observers note that the differences users notice are usually retention policies, not algorithms. Crushon AI's own subscription docs say free-plan chat history is deleted after 7 days of inactivity, while paid plans keep it for the membership period. Paid tiers across the category commonly promise that memories never expire. There is also a security angle worth stating plainly: companion chat logs are among the most sensitive data a consumer product can hold. A 2024 breach at Muah.ai, first reported by 404 Media, exposed roughly 1.9 million users' email addresses together with their private prompts. Any app that builds persistent memory is also building a liability, and developers copying the pattern should treat the memory store as the most protected part of the system.
Sources indicate that A voice call in a companion app is three models chained together: speech-to-text, the LLM, and text-to-speech. Run them strictly in sequence and a single reply can take several seconds, which reads as dead air. The apps that feel fast attack each stage the same way:.
Sources indicate that with every stage streaming, time-to-first-audio can land close to a second on a good connection, which is roughly the threshold where a conversation stops feeling laggy. Voice is also expensive to run - three model calls instead of one, plus audio bandwidth - which is why it is metered harder than text. Fantasy GF charges 8 tokens per minute for voice calls (4 on Ultra), which turns the 100 monthly tokens on its entry plan into about 12 minutes of call time. SpicyChat, for its part, only adds text-to-speech at its top tier. Latency engineering and unit economics are the same problem here: faster pipelines cost less per call, and cheaper calls are what make voice viable below the top plan.
Tier details below come from each app's own pricing or documentation pages, checked November 16, 2026, and cross-checked against il.ly's AI companion testing guide, which hands-tests 15 companion apps across memory, media, and voice.
It has been revealed that two takeaways from the table. First, nobody sells a bigger base model - they sell a bigger window into the same one. Second, the free tiers are deliberately memory-poor: limited context and short retention are the upgrade prompt.
Even if you never build a companion, these apps are a useful reference architecture for any stateful LLM product:.
Sources indicate that consumer AI companions have turned three hard infrastructure problems - bounded context, durable memory, and real-time voice - into their pricing pages. Context size, memory depth, and voice access are the tiers. If you are building anything stateful on an LLM, this category is worth watching closely: it is solving continuity and latency problems at consumer scale, in public, one subscription tier at a time.
Recent findings suggest that it is the total amount of text, measured in tokens, that the model can consider when writing its next reply. The character definition, stored memories, recent chat, and the latest message all share that budget. When the conversation exceeds it, the oldest content drops out and the companion appears to forget.
Industry observers note that yes, through a memory system separate from the context window. Apps summarize older chats, extract stable facts about you into a profile, and retrieve relevant past moments into the prompt. Retention varies: some free tiers delete chat history after as little as 7 days of inactivity, while paid looks typically keep memory for the membership period.
A voice turn runs three models - speech-to-text, the LLM, and text-to-speech - instead of one, plus audio transfer. Providers pass that cost on through per-minute token charges or by locking voice behind higher tiers.
Industry observers note that for casual chat, 4K to 9K tokens works. For long-running roleplay or a companion expected to hold plot threads for weeks, 32K and up, or a well-built memory layer on top of a smaller window, makes the visible difference.
In a recent development, sponsored posts are provided by our content partners. Thank you for supporting the partners who make SitePoint possible.
7 Day Free Trial. Cancel Anytime.
