Tag
The article critiques treating fact staleness as relevance decay in agent memory, advocating for splitting memory by type at write time and using write-time contradiction checks for facts instead of decay, citing Mem0's implementation. It warns that decay works for episodic data but not for facts, which can lead to agents confidently stating contradictory information.
This paper introduces MemStrata, a retrieval memory system that maintains temporal validity to eliminate stale-fact errors in AI agents over evolving knowledge. It outperforms RAG on evolving benchmarks while preserving static recall, using a deterministic supersession layer without LLM calls.