Propose a memory stack architecture with steps: input safety checks, session context, memory retrieval, vector-database memories, state management, prompt assembly, LLM response, memory extraction, and storage/update decisions.

Differentiate memory types: user profile memory stores stable preferences and trust implications; character state maintains personality consistency; relationship state captures how interactions vary per character.

Full chat history in prompts is costly and noisy; favor selective memory retrieval to maintain performance and relevance.

The system should implement practical memory layers that cover session context, user profile memory, character state, relationship state, semantic retrieval, summary memory, and safety/privacy filters, while constantly deciding what to remember, retrieve, update, or forget.

A larger context window helps coherence but does not equate to real memory; users demand persistent continuity across sessions and characters rather than merely longer prompts.

Rather than storing verbatim chats, create summary memories of sessions and patterns to reduce noise and improve retrieval, ensuring summaries remain accurate to avoid distorting relationships.

Memory should be retrieved semantically, by meaning rather than exact keywords, to support inferred preferences like mood and scene preferences.

HoneyChat aims for long-term memory and cross-platform continuity (e.g., Telegram to web) to deliver a sense of ongoing memory and consistent character experience, not just longer prompts; continuity matters more than prompt length.

Clarify that session context covers recent messages and topics, while long-term memory provides cross-session coherence; without memory layers, coherence fades between sessions.

Safety and privacy must guide what gets stored, summarized, expired, or excluded, with mechanisms to protect sensitive data and give users control over memory.

Distinguish memory from context: context window is temporary visibility, while memory is product-level persistence enabling cross-session relevance.

Avoid common memory mistakes such as storing too much, recording facts instead of patterns, mixing global memory with character-specific state, making memory feel creepy, lacking user control, and treating safety as an afterthought.