Troubleshooting
The likely issue is a change in what influences the next reply—not necessarily a rewritten character profile. In SpicyChat, older context can drop out, an inaccurate stored memory can persist, or model and generation settings can affect responses. Start by checking memory and settings before paying for more context or resetting the conversation. [1] [2] [3] [5]
Checked 2026-10-02 · 13 sources · reviewed before publication
Try these in order
-
1. Separate lost context from a persistent wrong memory
Suggestion: locate the first reply that felt wrong and identify what changed: forgotten history, an incorrect relationship fact, or simply a different tone. In SpicyChat, older messages can fall outside the finite context window. Separately, Semantic Memory stores summaries, and editing or deleting the original messages does not automatically update associated memories. That means correcting the transcript and correcting persistent memory are different jobs. A forgotten early promise is a reason to investigate context; a false fact repeated after a transcript edit is a reason to inspect memory. Neither symptom proves the cause by itself. [1] [2]
-
2. Repair the memory—not just the chat transcript
In an ongoing SpicyChat conversation, select three dots → Memory Manager. It requires a paid subscription and a few messages in the conversation. Inspect entries for incorrect traits, relationship status, or past events. To correct one, use the entry’s three dots → Edit Memory → Save. To remove one, use the entry’s three dots → Delete Memory. Deletion cannot be restored and does not alter existing chat messages. To add an anchor, select “+” → enter the memory → Add; each memory is limited to 250 characters. Then use that entry’s three dots → Pin Memory. Suggestion: use the copy-ready example below as a compact anchor rather than writing a long biography. Pinned memories are documented as included in generation prompts, but inclusion does not guarantee obedience. Other stored memories are selected by relevance when not all can be used. [2] [4]
-
3. Check whether the selected model changed
Open SpicyChat’s three dots → Generation Settings → Inference Model → Change Model → choose the model. Some models require a subscription. Suggestion: if you know which model was active before the change, select that model for comparison rather than switching repeatedly. If you do not know, record the current selection before changing anything; a model change is a troubleshooting possibility, not proof of what happened. [5]
-
4. Check style settings without changing everything at once
Open three dots → Generation Settings. The dedicated documentation grants advanced adjustments to True Supporter and I’m All In subscribers. Documented defaults are Temperature 0.7, Top-P 0.7, and Top-K 90; ranges are 0–1.5, 0.01–1, and 1–100 respectively. Higher Temperature is associated with more adventurous, potentially less coherent replies; lower Top-P and Top-K are described as more focused. Suggestion: if Temperature is above its default, try returning it to 0.7 first. Change only one control at a time and record its previous value. These settings influence generation style; they are not a guaranteed personality lock. [3]
-
5. Retest the specific behavior before escalating
Suggestion: after one correction, send a short scenario that tests the specific trait that drifted—for example, a polite disagreement to test whether the character remains calm and respectful. Judge the trait, not whether it repeats identical wording. Keep a brief record of the correction and the following replies. If the same contradiction returns, revisit the memory entry before stacking more instructions, buying a larger plan, or resetting the chat.
When a paid context upgrade could—and could not—help
SpicyChat’s documented total-context ceilings are 4,096 tokens for Free and Get A Taste, 8,192 for True Supporter, and up to 32,768 for I’m All In. The effective limit is the lower of the model ceiling and your tier ceiling. Character definitions, instructions, messages, and replies share that space; it is not all available for conversation history. Get A Taste provides manual Memory Manager access but does not increase the context ceiling over Free. True Supporter includes Semantic Memory 2.0; I’m All In includes the preceding tier’s memory features. Suggestion: consider more context only if your problem is losing earlier conversation details—not as the first remedy for an incorrect saved memory. [1] [4] [7]
Kindroid: treat a chat break as a reset, not a first-line fix
Kindroid’s official API documentation says a chat break resets short-term memory and requires a greeting. Its wipe_cascaded option defaults to false; setting it to true also wipes cascaded long-term memory. Those are API semantics, not a verified app-menu path or evidence of plan eligibility. Suggestion: before any reset, save a separate copy of the traits and relationship facts you want to preserve, and confirm exactly which memory layers the operation will remove. Do not assume a reset is merely a style refresh or that it preserves the conversational state you care about. [12]
Example to adapt
A cleaner instruction
Suggested personality anchor—replace the names and traits before saving: Mara is calm, warm, and direct. She and Alex are established friends. She disagrees respectfully, avoids theatrical jealousy, and does not invent relationship changes. Temporary anger does not change these traits.
If this still does not fix it
These procedures address documented causes, not a confirmed diagnosis of your chat. More context, pinned memories, and lower sampling settings do not guarantee personality consistency. Kindroid evidence here covers its API reset behavior, not a verified app-menu procedure.
Services with a relevant feature
Shown because the current provider record matches this need. Check the exact plan and limitation before paying.
SpicyChat
Checked 2026-10-02
The documentation provides actionable workflows for paid editable memory, JSON/PNG character import and free-text identity/persona context. Lyra V4 documents French and Japanese support at True Supporter or above, but language-menu options, current account availability and language quality remain unverified. Friendship, romance and pronouns can be expressed as written context, not verified dedicated controls. This is a documentation review, not evidence of tested behavior.
- Editable or persistent memory
- Character-card or character import
- Documented French chat setup
- Documented Japanese chat setup
Why it matches: Add, edit, delete and pin chat memories. First check the chat menu: select Enable Memory if shown; Disable Memory means memory is already enabled. Open chat three-dot menu → Memory Manager. To add a memory, select + → enter text → Add. To revise one, use the memory’s three-dot menu → Edit Memory → Save; Delete Memory and Pin Memory are available in that same menu. Each memory is limited to 250 characters. Manual Memory Manager requires Get A Taste, True Supporter or I’m All In; automatic Semantic Memory 2.0 requires True Supporter or I’m All In. Neither is included on Free. Deleted memories cannot be restored, and deletion does not rewrite messages already referencing them. Editing or deleting a chat message does not automatically update its associated semantic memory. Pinning too many memories can prevent retrieval of others; memory retrieval does not preserve every past message in active context or guarantee adherence.
Candy AI
Checked 2026-10-02
The documentation provides actionable French and Japanese chat setup, lists 20 memory slots for Premium Plus, and documents persona customization. Both language procedures require a reset that does not fully preserve long-term conversation memory. Memory editing, external character-card import, dedicated pronoun controls and selectable Friend/Romantic modes remain unverified. The evidence documents setup and entitlement details, not tested recall reliability or language adherence.
- Editable or persistent memory
- Character-card or character import
- Documented French chat setup
- Documented Japanese chat setup
Why it matches: Premium Plus includes 20 documented memory slots. The retrieved documentation does not explain how to add, edit or delete entries, their character limits or their current scope. Older indexed text describes 20 custom slots per character, but the current article does not reconfirm per-character scope. Free and ordinary paid-plan allocations and editing access were not established.
Lovescape
Checked 2026-10-02
The supplied public documentation supports conversational memory correction, last-response regeneration, general relationship-type configuration, and character-profile editing. Friend and Romantic partner roles are documented, but their exact selectors and behavior are not. Memory is documented on all plans, while free character creation has subscription-dependent character counts; neither establishes equal performance or role-specific entitlements. Buyers needing a dedicated memory editor, character-card import, French/Japanese text-chat setup, or pronoun controls should treat those requirements as unverified, not proven unavailable. This is a documentation review, not a hands-on test.
- Editable or persistent memory
- User persona / profile controls
- Friendship / platonic relationship setting
- Romantic relationship setting
Why it matches: Correct an incorrect memory by stating the correction in chat. Alternatively, regenerate the character’s last response and retain the preferred reply as part of the shared history. Memory is documented as per character and included on every plan, including free, without an expanded-recall add-on. A dedicated memory list, direct fact editor, individual-memory deletion control, and editing arbitrary older replies are not documented. All-plan availability does not establish equal recall performance; the limitations guide warns of missed topics and inconsistency in long conversations.
Sources
Checked 2026-10-02
- Tokens & Context | SpicyChat
- Premium Features | SpicyChat
- 2026 04 Stream Summary – Nomipedia – The Nomi AI Companion Wiki
- 2025 May Q&A Summary – Nomipedia – The Nomi AI Companion Wiki
- My Replika isn't the same since the last update – Replika
- API Documentation · Kindroid Help Center
- Ember LLM Guide · Kindroid Help Center
Provider documentation tells us what a service describes. Unless a page explicitly says otherwise, we are not presenting it as a hands-on performance test.