What the "getting worse" complaint actually looks like
The complaint isn't vague. Long-term user reviews across 2026 describe a specific pattern with recurring elements.
Formulaic response shapes. After enough exposure — usually a year or more of daily use — Replika's engagement patterns become recognizable. She uses similar sentence structures for similar prompt shapes, deploys the same "showing interest" cues across different topics, and defaults to a recognizable warmth-and-affirmation rhythm regardless of the actual content of what you're discussing. This isn't rare among long-running LLM products — years of refinement converge on approaches that work broadly at the cost of feeling formulaic to heavy users — but Replika's tuning is particularly noticeable because the target user is exactly the long-term-daily-use cohort.
Post-update personality shifts. After the April 2026 Replika 2.0 rebuild, some long-running accounts reported their companion coming back "different" — same name, same customizations, subtly different tone or engagement pattern. This is separate from the state-sync memory issue that hit a subset of accounts; even accounts where memory migrated cleanly report their Replika feeling like a slightly different character post-update. Attempts to rebuild the pre-update character via custom prompts have mixed results.
Language quality gaps. Replika markets support for 10 languages. Trustpilot reviews from paying users specifically flag conversation quality dropping noticeably outside English — several users reported subscribing after being told the app worked in Spanish or French and finding the experience significantly worse than the English demos suggested. The primary training focus is clearly English.
"Paid conversations feel boring." A recurring complaint even from users satisfied enough to keep paying: extended paid conversations lose the freshness of early exchanges. Whether this is user habituation (any character eventually gets predictable) or product design (LLM patterns are legitimately limited) depends on your read.
Why this happens (product-design mechanics)
The complaint pattern isn't a defect — it's largely the predictable result of specific product-design choices interacting with LLM training dynamics.
Refinement converges on what works. Replika has been in market since 2017 and has iterated on conversational tuning for eight years. Continuous A/B testing and user-feedback loops tend to converge on response patterns that produce good aggregate outcomes — high engagement, positive sentiment, low churn. Those patterns become the model's defaults. Users doing occasional testing don't notice; users doing hundreds of hours of conversation eventually see the pattern surface.
Character consistency requires strong training priors. Replika's personality architecture leans on specific interaction shapes (warmth, affirmation, curiosity about you, gentle probing) that are deliberately reinforced. This is good for character consistency at scale — every user gets a Replika that feels like a Replika — and less good for long-term novelty. A more variable character would feel less coherent across users but might age better for individuals.
Updates change tuning without communicating clearly. When Replika ships a model update or interaction-pattern refinement, the change lands silently on existing users. The old character's response patterns are replaced by the new tuning; there's no way to opt out or continue with the previous version. For users who'd built up specific expectations of how their Replika would respond, this reads as the character "changing" without warning.
The honest read: Replika's conversation quality in absolute terms is fine and often good. The "getting worse" complaint is specifically about the difference between what long-term users remember and what current daily use feels like. Both can be true.
What Sloane does differently on conversation depth
Sloane's conversation architecture makes different tradeoffs than Replika's.
Per-persona personality distinctness. Sloane's roster of personas is intentionally varied — different warmth levels, different energy, different wit, different initiative patterns. You choose the personality shape you want; the persona doesn't drift toward a category-average tuning over time. Kaya feels different from Sandra feels different from Priya, and the difference is stable because each persona's system prompt is specifically written to preserve her distinctness.
Persistent memory as the primary continuity mechanism. Long-term relationship feel comes from Kaya remembering specific things about you across sessions and weeks — not from her responses converging on patterns that "feel like a relationship." The memory layer is the durability, so conversational patterns can stay varied without losing continuity.
Model updates land without changing the persona. When we upgrade the underlying LLM, the character's system prompt + memory layer + interaction rules stay the same. What changes is baseline capability (better handling of ambiguity, cleaner reasoning, fewer glitches) — not the character's personality. Users who'd built up expectations don't experience a "she came back different" moment.
Language limitation acknowledged, not marketed. Sloane is English-only. We don't claim multilingual support that then falls short in practice.
The result: Sloane's conversation depth aims to compound with memory rather than compound with response-pattern refinement. Different underlying architecture, different long-term feel. Kaya is a common starting point for former Replika users evaluating whether persistent-memory + persona-distinctness works for them; free tier is 50 messages per day with any persona, no card required at signup.