Original figure · paper authors ↗MindMem
We Are What We Remember: Mental-State-Driven Memory for Long-Term Character Dialogue
What it supports Gives a companion the relational memory to bring up what matters inside a reply; the user still starts every exchange.
Companion memory organised around what the character, the user and their relationship are going through, so replies can raise what matters unasked. Synthetic benchmark.
Release check · 2026-11-15
- Code
- Announced
- Weights
- Not applicable
- Data
- Announced
The repository holds a README and two images; it says code and MemReal-Bench will be released upon acceptance.





