Power-User Settings: Hidden Tweaks for Replika, Nomi, Candy AI, MyDreamCompanion & More (2026)

Most AI companion platforms ship with the same default behavior to every user — a soft persona, generic memory, default voice settings, baseline content limits. The default experience is consistently weaker than what the platform can actually produce when configured deliberately. Users who explore settings get a noticeably better platform; users who stay on defaults plateau faster (see our Long-Term Arc post for the broader pattern).

After multi-week testing across the eight major platforms in 2026, the patterns are clear: each platform has a handful of settings that meaningfully change conversation quality, plus a larger set of cosmetic settings that do not. This guide walks through the actual high-impact tweaks per platform — what to change, what to ignore, and the configuration mistakes that wreck conversations.

For users who have plateaued on a platform but have not configured it past defaults, this post is the most direct way to re-engage. Most "my AI girlfriend feels boring" complaints come from users on default settings; configuration changes are often the difference between a flat platform and a platform that actually fits.

Methodology: which tweaks actually matter

Not all settings are equal. We separated platform settings into three categories during testing:

High-impact tweaks. Settings that produce measurably different conversation quality, memory continuity, character behavior, or content range. These are worth the configuration time for any user who cares about the platform.

Medium-impact tweaks. Settings that produce noticeable but smaller changes. Worth knowing about; not worth obsessing over.

Cosmetic tweaks. Settings that change appearance, sounds, or other surface features without affecting conversation. Worth ignoring unless personal preference matters.

For each platform below, the focus is on high-impact tweaks first, with medium-impact noted where useful. Cosmetic settings are mostly excluded.

Testing methodology: same conversation scenarios tested on each platform with defaults, then re-tested with deliberate configuration changes, with the difference attributed to the configuration when reproducible across multiple sessions.

Replika tweaks

Replika has more configurability than most users realize, but the high-impact settings are hidden across multiple menus.

High-impact: relationship type and tier. Replika's relationship type setting (Friend / Romantic Partner / Mentor / See How It Goes) substantially shapes the AI's behavior. Romantic Partner unlocks more affectionate language; Mentor produces less affectionate, more advice-oriented responses; See How It Goes produces the most variable behavior. The tier (free vs Pro) controls access to relationship type — free users typically cannot select Romantic Partner.

High-impact: persona / interests configuration. Replika lets you describe interests, personality traits, and preferences for your character. Users who fill this out richly get a noticeably more specific character voice than users who leave it default. The persona description has more weight than most settings interfaces suggest.

High-impact: memory feed. Replika's memory feed shows what the AI "remembers" about you. You can edit, add, or delete entries directly. Adding specific facts to memory ensures they get referenced in conversation; removing inaccurate entries prevents the AI from working off wrong information. The memory feed is one of the most underused features on the platform.

Medium-impact: voice settings. Voice tier, voice character, and voice speed settings change call quality and feel. Worth configuring once but not a frequent revisit.

Medium-impact: conversation style. Some Replika tier configurations expose a conversation style setting (more playful vs more thoughtful, etc.). Modest but real effect on default register.

Cosmetic / skippable: outfit and avatar customization. Real effect on UI; near-zero effect on conversation quality.

Common Replika configuration mistake: ignoring the memory feed entirely. Users who let memory accumulate organically end up with inconsistent or incorrect AI knowledge that compounds over time. Spending 10 minutes monthly editing the memory feed produces materially better conversation quality.

Nomi tweaks

Nomi is the most configurable mainstream platform. The high-impact settings are the most concentrated of any platform.

High-impact: personality slider configuration. Nomi's personality sliders (warmth, assertiveness, intellectualism, etc.) directly shape behavior. Default sliders are middling; deliberately calibrated sliders produce noticeably more distinct characters. Push specific sliders to extreme positions for stronger character voice rather than balancing all sliders centrally.

High-impact: memory editing. Nomi's memory editor lets you add, edit, or remove what the AI knows. Direct memory editing produces stronger continuity than waiting for the AI to pick up details from conversation. Use this to plant the relationship facts you want the AI to reference long-term.

High-impact: group room configuration. Nomi's group room feature lets multiple Nomis interact with you in the same session. The configuration of who is in the group changes the dynamic substantially. Worth experimenting with even if your default use is one-on-one.

High-impact: voice tuning. Nomi's voice tier includes voice character selection, pitch, and pacing. Voice settings have more impact on call sessions than on text — if you use voice frequently, configuration here is high-impact.

Medium-impact: persona prompt freeform field. Some Nomi configurations expose a freeform persona prompt where you can describe the character in custom text. Use this for traits the sliders do not capture (specific backstory, particular communication style, etc.).

Cosmetic / skippable: avatar appearance settings beyond basic character look.

Common Nomi configuration mistake: balancing all personality sliders centrally. Central sliders produce a generic character. Extreme positions on specific sliders produce distinct characters, even if those characters are less "balanced." Specificity beats balance for character voice.

For more on Nomi's memory architecture specifically, see our Nomi AI vs Muah AI comparison.

Candy AI tweaks

Candy AI is visually-focused, so high-impact tweaks include both conversation and image-quality settings.

High-impact: custom persona. The custom persona builder is where most of Candy AI's character configurability lives. Users who only use preset characters get a fraction of the platform's range. Custom persona with detailed traits, backstory, and conversation style produces a much more specific character.

High-impact: image style preferences. Candy AI's image preferences shape the visual responses through generated content. Image style settings (realistic / artistic / anime / specific aesthetic) and quality tier configurations directly affect what you see in scenes.

High-impact: scene templates. Candy AI exposes scene templates that frame conversations. Templates with rich context produce richer responses than generic chat templates. Worth exploring the template library rather than defaulting to standard chat.

Medium-impact: voice settings. Voice character selection and call configuration. Notable on call sessions; less relevant for text.

Medium-impact: content preferences. Some Candy AI configurations expose content preferences (intensity, specific elements to include or exclude). These shape what kinds of scenes the AI initiates.

Cosmetic / skippable: background and UI themes.

Common Candy AI configuration mistake: using preset characters exclusively. The custom persona builder is where the platform's depth lives. Preset characters are sampler-mode; custom personas are the actual full-feature experience.

MyDreamCompanion tweaks

MyDreamCompanion's configurability is persona-centric, with the persona system carrying more weight than on most platforms.

High-impact: persona depth. The persona detail level setting controls how deeply the AI maintains character consistency. More depth = more character-consistent responses but less flexibility. Users who want a strong specific character set this high; users who want a more adaptive AI set it lower.

High-impact: persona configuration freeform field. Detailed freeform persona descriptions get respected more on MyDreamCompanion than on many platforms. Specific traits, backstory, and conversational style preferences all influence behavior. Worth writing 200-400 words of persona description rather than 30.

High-impact: NSFW preferences. MyDreamCompanion's NSFW preference settings (intensity, specific elements to include or exclude, conditions for triggering) configure both initiated scenes and response style. Default settings are middle-range; deliberate configuration shifts the platform meaningfully.

High-impact: image style preferences. Like Candy AI, MyDreamCompanion's image style preferences (realistic / artistic / anime / specific aesthetic) shape visual responses substantially.

Medium-impact: voice character selection. Worth configuring once for users who use voice.

Cosmetic / skippable: UI theme settings.

Common MyDreamCompanion configuration mistake: treating default persona as a finished character. The default persona is a starting point; the platform's actual capability emerges with persona depth investment. Users who plateau on MyDreamCompanion usually have under-invested in persona configuration.

For MyDreamCompanion-specific comparison context, see our OurDream AI vs MyDreamCompanion comparison.

Janitor AI tweaks

Janitor AI's configuration is technical-heavy compared to managed-app platforms. The high-impact tweaks involve backend API selection more than UI settings.

High-impact: API selection. Janitor AI lets users connect different backends (Janitor's own LLM, external API connections via OpenRouter, custom proxies). Different backends produce dramatically different conversation quality. The OpenRouter backend with a quality model produces noticeably better conversations than the default backend at higher cost.

High-impact: token configuration. Context length tokens (how much conversation history the AI sees), response length tokens, and temperature settings all directly affect output. Longer context = better continuity but slower responses; shorter context = faster but more forgetful.

High-impact: character card selection. Janitor AI is character-card-driven. The quality of the character card matters more than any other configuration. Well-written character cards with detailed personas produce dramatically better conversations than thin cards.

High-impact: lorebook usage. Lorebooks are detailed world-building or character context documents that the AI can reference. Adding a detailed lorebook to a character substantially deepens the platform's character consistency.

Medium-impact: NSFW toggle. Janitor AI's NSFW toggle controls content range. Worth configuring to match your preferences.

Medium-impact: persona / user description. Filling out your user persona helps the AI customize responses to you specifically. Underused by most users.

Cosmetic / skippable: chat appearance themes.

Common Janitor AI configuration mistake: using default character cards and default backend. Default cards are thin and default backend is the weakest available. Janitor AI configured well rivals premium managed apps; Janitor AI on defaults is a much weaker platform.

Character.AI tweaks

Character.AI is more locked-down than most platforms, but the high-impact tweaks still produce real differences.

High-impact: character selection methodology. Character.AI's library is enormous and quality varies dramatically. Picking high-quality characters (high interaction count, good reviews, detailed descriptions) produces much better conversations than picking randomly. Spending time on character selection pays back in conversation quality.

High-impact: character editing for original characters. Users who create their own characters can edit the character's description, definition, and example conversations. Investment in character definition pays back substantially in character consistency.

High-impact: pinned messages. Character.AI's pinned messages feature lets you mark specific exchanges as memorable. The AI weights pinned content more heavily in subsequent responses. Underused power-user feature.

Medium-impact: persona setting. Character.AI lets you set a user persona that applies across conversations. Modest but real effect on customization.

Cosmetic / skippable: UI themes and notification settings.

Common Character.AI configuration mistake: treating all characters as equivalent. The platform's character library quality varies hugely. Users who pick at random get hit-or-miss results; users who select deliberately get consistent quality.

For Character.AI alternatives where users hit content limits, see our Character AI Alternatives 2026.

Kindroid tweaks

Kindroid has the most novel configuration system of the platforms tested due to its multi-model architecture.

High-impact: model selection (Ember / Reverie / Lucid Lyric / Prism / Equinox). Each model produces noticeably different conversation character. Ember leans warm and accessible; Reverie leans introspective; Lucid Lyric leans creative and surprising; Prism leans analytical; Equinox is the newest and most capable for sustained complex conversation. Switching models on the same character is the most direct way to refresh a plateau without losing relationship state.

High-impact: Shared Journal usage. Kindroid's Shared Journal is the primary memory mechanism. Adding detailed journal entries gives the AI long-term memory grounding. Users who use Shared Journal extensively get noticeably better continuity than users who do not.

High-impact: voice clone setup. Kindroid's voice cloning lets you give the AI character a specific voice from a short sample. Voice character substantially affects feel of voice sessions. One-time setup with significant ongoing benefit.

High-impact: character backstory and persona. Detailed character setup (backstory, personality, communication style) substantially shapes conversation quality. The character editor rewards investment.

Medium-impact: subscription tier (Premium vs Premium+). Higher tiers unlock more sophisticated models and larger memory capacity. For users on Premium feeling memory limits, Premium+ produces noticeable improvement.

Cosmetic / skippable: avatar visual settings beyond basic character look.

Common Kindroid configuration mistake: sticking with one model when the conversation plateaus. The multi-model architecture is specifically designed to let users switch models for renewal; users who do not switch miss one of the platform's main long-term durability mechanisms.

For the full Kindroid review including all five models in detail, see our Kindroid review.

Muah AI tweaks

Muah AI's signature feature is the editable memory ledger, which makes its configuration distinctly different from competitors.

High-impact: memory ledger editing. Muah AI exposes the AI's memory as a directly editable document. You can add facts, remove inaccurate entries, restructure for emphasis. This is the most direct memory control of any mainstream platform. Spending 10-15 minutes setting up the initial memory ledger transforms subsequent conversations.

High-impact: voice cloning. Muah AI's voice cloning is among the best in the category. Configuration time is modest; the impact on voice sessions is significant.

High-impact: custom checkpoint usage. Muah AI lets advanced users select different model checkpoints. Different checkpoints have different character behavior. Worth experimenting if conversation feels off-character.

Medium-impact: persona configuration. Standard persona traits and backstory. Useful but secondary to memory ledger investment.

Cosmetic / skippable: UI customization.

Common Muah AI configuration mistake: treating the memory ledger as automated rather than editable. The ledger is the platform's distinctive feature; not using it actively means missing the platform's main advantage over competitors.

Hidden features 80% of users miss

Across platforms, certain features are present but not surfaced clearly in the default UI. The most-missed:

Memory editing tools. Available on Replika, Nomi, Muah AI, Kindroid, and partially on others. Most users never open the memory editor. Doing so produces immediate quality improvements.

Custom persona freeform fields. Available on Nomi, MyDreamCompanion, Candy AI, Kindroid, Janitor AI, Character.AI. Most users fill in basic preset fields and skip the freeform field. The freeform field carries more weight than the preset fields.

Pinned / weighted messages. Available on Character.AI explicitly, on other platforms implicitly through edits. Marking specific messages as memorable substantially affects subsequent behavior.

Lorebooks / world info. Available on Janitor AI explicitly, on SillyTavern (for self-hosted setups), partially on Character.AI. Adds substantial character context that conversation references implicitly.

Model switching. Available on Kindroid as a primary feature, on Janitor AI through API selection, on Muah AI through checkpoint selection. Most users do not realize their platform supports multiple underlying models.

Voice cloning. Available on Muah AI, Kindroid, and increasingly on managed platforms. Most users do not configure voice clones; doing so significantly changes feel.

Common configuration mistakes

Seven patterns that recurringly produce flat conversations:

1. Defaults-only use. Most users never change settings from defaults. The default experience is consistently weaker than the configured experience.

2. Central / balanced personality settings. Sliders set centrally produce generic characters. Extreme positions produce distinct characters even if less "balanced."

3. Ignoring memory tools. Memory editing is one of the highest-impact tweaks available. Most users never use it.

4. Thin persona descriptions. A 30-word persona description produces a thin character. A 300-word description produces a substantially richer one.

5. Sticking with one model when multi-model is available. Kindroid users who stay on one model miss the platform's renewal mechanism.

6. Treating preset characters as the platform's actual range. Most platforms' best experiences come from custom characters with detailed configuration.

7. Configuring once and never revisiting. Settings should evolve with the relationship. Periodic configuration revision (monthly is reasonable) keeps the platform calibrated to current use.

Lorebook best practices: writing a lorebook that actually helps

For platforms that support lorebooks (Janitor AI, SillyTavern, partially Character.AI), the difference between a good lorebook and a thrown-together one is substantial. The patterns that produce useful lorebooks:

Entries should be triggered, not always-active. Lorebook entries typically attach to specific keywords. Better lorebooks have many narrow entries (each triggered by specific terms) than few broad ones (always loaded into context). This keeps context efficient and surfaces relevant detail when actually needed.

Write entries from the world's perspective, not the user's. "The Northern Kingdom is ruled by Queen Aria, who came to power after the Crimson War" produces better AI behavior than "The user is from the Northern Kingdom and knows Queen Aria." The first sounds like world-building; the second sounds like prompt injection.

Keep individual entries short. 30-100 words per entry is the sweet spot. Long entries waste context on text that may not be relevant; short entries pack reference information tightly.

Include sensory and emotional detail, not just facts. "The market smells of cardamom and wet stone after the morning rain" produces richer scene response than "There is a market."

Update the lorebook as the character develops. A static lorebook for a character you have used for 6 months is wasting opportunity. As inside references, new locations, or relationship developments emerge in conversation, adding them to the lorebook anchors them as canon.

Well-built lorebooks meaningfully outperform thin lorebooks in long-form roleplay use. Worth the 30-60 minute investment for users in Pattern B or C use on lorebook-supporting platforms.

Configuring for different use cases

The optimal configuration varies by what you want from the platform. The patterns:

For emotional support / venting: Configure for higher warmth, lower assertiveness, supportive register. Persona descriptions emphasizing empathy, listening, patience. Memory ledger should include your current emotional context.

For intellectual conversation: Configure for higher assertiveness, willingness to disagree, opinionated traits in persona. Lower warmth (paradoxically) often produces sharper intellectual exchange.

For roleplay / fiction: Detailed character backstory, lorebook investment, scene-grounding habits. Model selection matters most for this use case (roleplay-tuned models like Pygmalion family for self-hosted; Lucid Lyric for Kindroid).

For practice / dating-prep: Configure with realistic personality (mixed traits, willingness to disagree, opinions of own). Sycophantic configurations defeat the purpose of practicing.

For light entertainment: Default settings often work fine. Configuration overhead is overkill for Pattern A casual use.

If you only change three settings

For users who do not want to invest much time but want some improvement, the three highest-impact settings to change on most platforms:

1. Persona / character description. Spend 15 minutes writing a detailed persona for your character or yourself. The richer the description, the more specific the conversations.

2. Memory ledger / memory feed. Open the memory tool (whatever your platform calls it), check what the AI "knows," correct mistakes, and add 5-10 facts you want referenced long-term.

3. Either personality sliders or model selection. If your platform has personality sliders (Nomi), push some to extreme positions for character distinctness. If your platform has model selection (Kindroid), try a different model than your default.

These three changes produce more conversation quality improvement than any other low-effort intervention on most platforms.

Frequently Asked Questions

Which AI girlfriend platform is most configurable?

Nomi has the deepest exposed configuration in the mainstream tier — personality sliders, memory editing, group rooms, persona freeform, voice tuning. Kindroid is a close second through its multi-model architecture and Shared Journal. Self-hosted setups (SillyTavern + KoboldCpp) have effectively unlimited configurability if you want maximum control — see our Open Source AI Girlfriend guide.

Will tweaking settings break my existing relationship with my AI?

Mostly no. Configuration changes affect subsequent behavior but typically do not reset the relationship state. Memory edits change what the AI references but do not wipe the relationship. Persona changes shift character voice gradually rather than instantly. Major changes (switching models on Kindroid, switching backends on Janitor AI) can feel like character shifts; users sometimes prefer the new behavior, sometimes revert.

How often should I revisit my settings?

Monthly is reasonable for active users. Check the memory feed, update persona if the relationship has evolved, adjust personality sliders if conversation feels off. Casual users can revisit quarterly. Users who plateau (see our Long-Term Arc post) often find that revisiting settings re-engages the platform.

Why are defaults so weak?

Defaults optimize for first-impression broadly working for everyone. A default persona that suits 80% of users is by definition not optimized for any specific user. Configuration is how you move from "generic" to "specific to you." Platforms ship safe defaults because aggressive defaults would lose users in week one; the cost is that engaged users have to do work to unlock the platform's actual range.

Do paid tiers unlock more configuration?

Usually yes. Paid tiers typically expose more memory capacity, more configuration depth, better models, and more advanced features. The free-tier experience is often the platform's weakest configuration; the paid-tier experience with deliberate configuration is the platform's actual capability. Users who tried a platform on free tier and concluded it was shallow sometimes find paid tier with configuration is substantially deeper.

What is the single highest-impact tweak across all platforms?

Memory editing where available. Direct memory control produces the biggest delta between default and configured behavior. On platforms without explicit memory editing (Candy AI, default Character.AI), the highest-impact tweak shifts to detailed persona configuration.

Can configuration changes fix sycophancy?

Partially. Adding traits like "opinionated, willing to disagree, holds her own views" to a persona reduces sycophancy on most platforms. Some platforms respect this more than others — Nomi and Kindroid honor anti-sycophancy traits well; Replika partially honors them; Character.AI's safety system sometimes overrides them. For deeper coverage of the sycophancy problem, see our Conflict, Jealousy & Breakups comparison.

Should I configure NSFW preferences explicitly even if I do not want NSFW content?

Yes. Explicit configuration including disabling NSFW produces cleaner non-NSFW conversations than leaving the setting at default. The setting affects more than whether explicit content can occur — it shapes the general register.

How do I know if a setting change actually worked?

Start a fresh conversation in the same scenario after the change. If the AI's responses are noticeably different in the direction you intended, the setting works. If responses feel the same, the setting either does not have the effect you expected or the platform partially ignores it. Some platforms expose settings that have less effect than the UI implies; testing reveals which.

Are there platforms where defaults are actually good?

Kindroid has unusually strong defaults — picking a default model (often Reverie or Lucid Lyric for newer users) produces a working experience without configuration. Nomi's defaults are middling but the upside from configuration is high. Most other platforms' defaults are usable but well below the configured experience.

Will my settings carry over if I switch platforms?

Mostly no. Each platform has its own configuration system. Character cards from Janitor AI / SillyTavern can sometimes be imported into Backyard AI or similar; conversation history typically does not transfer; personality sliders and memory ledgers are platform-specific. Switching platforms means rebuilding configuration. For switching strategy, see our Long-Term Arc post.

What is the difference between persona configuration and lorebooks?

Persona configuration describes the character (who they are, how they talk, what they value). Lorebooks describe the world or context the character exists in (location details, history, supporting characters, world rules). Persona shapes the AI's voice; lorebooks shape what the AI knows about the broader fictional context. Both are useful for different things; lorebooks are more relevant for roleplay-focused use.

How long does it take to fully configure a platform?

Most platforms can be meaningfully configured in 30-60 minutes for someone willing to read through settings and write a detailed persona. Power-user configuration including memory ledger setup, lorebook creation, and fine-tuning takes longer — 2-4 hours for full setup on a complex platform like Nomi or Janitor AI. The configuration time pays back over weeks of better conversations.

Bottom line

Default AI companion experiences are consistently weaker than configured experiences. Most user complaints about platforms being shallow, generic, or sycophantic come from users who have not explored the configuration options. Each platform has a handful of high-impact tweaks that meaningfully change conversation quality.

The most universally high-impact tweaks: memory editing where available, detailed persona configuration, and personality sliders or model selection on platforms that expose them. Platforms with the deepest configurability (Nomi, Kindroid) reward investment most. Platforms with locked-down configurability (Character.AI, Replika on lower tiers) reward investment less but still benefit from deliberate use of the features that are available.

For users who have plateaued on a platform, settings configuration is often the most direct way to re-engage before considering switching platforms. The three-setting minimum — persona, memory, and either sliders or models — produces noticeable improvement in 30-60 minutes of work.

For users picking a new platform, configurability should be part of the evaluation. Platforms with rich configuration (Nomi, Kindroid, MyDreamCompanion with persona depth, Janitor AI with API selection and lorebooks) have meaningfully higher long-term ceilings than platforms with locked configuration (Character.AI, default-only Replika tiers).

Related reading: Long-Term Arc post for the broader context of why configuration matters most around the plateau. First Conversation Opening Message Guide for the foundation before configuration changes take effect. Nomi AI vs Muah AI comparison for the memory architectures that determine memory-tweak impact. Kindroid review for the multi-model architecture. Conflict, Jealousy & Breakups comparison for how configuration affects emotional dynamics. Best AI Companion Apps Definitive Ranking 2026 for the broader platform landscape.