Reproduce “CreaMem: A Scene-Aware Memory Architecture for Personalized Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qixuan Sun, Yue Que, Bowei He, Jin Guo, Dihang Yang, Wenchang Situ, Chen Ma
CreaMem organizes long-term agent memory around a user’s life scenes—Life, Work, and Interest—while pairing concrete episodic events with generalized scene-level traits. Across LoCoMo and LongMemEval-S, it reaches 54.61% and 66.40% GPT-4o-Judge accuracy, respectively, outperforming the strongest reported baselines while using far less context than full-history retrieval.
Key insights
- Partitioning memory by life scene reduces competition between unrelated memories, but excessive fragmentation can dilute retrieval density; the default three-scene organization outperforms a five-partition alternative.
- Combining episodic and scene-level representations is more effective than using either alone, improving access to both concrete evidence and generalized preferences.
- Retrieval quality is non-monotonic with context size: larger budgets can help initially, but excessive entries introduce noise and increase cost without improving answers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.