Chat memory

Kawaii Candy uses a hybrid memory system. As a creator, you control lorebook entries, starting situation, world information, and other character fields. During chat, the platform seeds live Current Context once, then updates it after each reply, and allows users to add manual overrides.

These layers work together—lorebook provides curated, keyword-triggered stable facts; starting situation bootstraps the opening scene; Current Context tracks live location, mood, and relationship state; custom memory lets the user pin notes the AI should always remember.

• Creator-controlled — lorebook (keyword-triggered stable facts), starting situation (one-time seed), world info, relationships, author's notes, jailbreak resistance
• Per-chat live context — Current Context merges world state and character memory; updated after each assistant reply and editable in chat settings
• Per-chat auto memory — rolling summary, pinned facts, timeline events, semantic recall of related past facts
• User override — custom memory textarea in chat settings (max 5,000 characters)

Auto memory is maintained after each assistant reply—not when the user sends a message. Extraction, summarization, and embedding run in the background after the character responds.

Users can view memory in chat settings under the Context tab.

The Context tab shows memory health stats (pinned count, timeline count, summary length, last updated) and structured memory panels.

• Current context — editable live scene fields: location, time of day, mood, active goal; plus read-only relationship and emotional state (saved on blur for world fields)
• Opening setup (reference) — read-only snapshot of the resolved starting situation used when this chat was created (legacy chats may still show placeholders until viewed with a persona)
• Pinned memories — read-only list of high-confidence facts with category and source badges; creator-authored pins can include {{variables}} and resolve when injected into the model context
• Timeline — read-only chronological events from the conversation
• Character memory details — read-only speech notes and shared experiences when present
• Conversation summary — read-only rolling summary when one exists

• Rebuild summary — regenerate the rolling summary from recent messages
• Reseed current context — rebuild live scene state from the opening setup snapshot and first opening messages
• Import custom memory — migrate legacy custom memory text into the new auto memory system
• Clear auto memory — remove extracted summary, pinned facts, timeline, character memory, and embeddings for this chat (does not clear lorebook or character fields)

Below the memory panels, an "Advanced override (custom memory)" accordion lets users type manual notes the AI should always remember. Maximum 5,000 characters. Supports {{variables}} like other chat settings fields.

Custom memory is always included in context alongside auto memory and character fields.

Lorebook entries are creator-defined and trigger on keywords in user messages—they do not change based on conversation history.

Auto memory is extracted per chat from what actually happened—promises made, preferences learned, relationship shifts, and plot beats.

Use lorebook for stable world facts ("The Silver District is lawless"). Auto memory handles dynamic story state ("The user promised to return tomorrow").

When vector search is available, the platform can retrieve semantically similar past facts from earlier in the conversation—not just keyword matches. If vector search is unavailable, retrieval falls back to keyword matching.

Creators do not configure this per character; it runs automatically as part of the memory system.