A memory may be outdated for present‑day reasoning yet indispensable for reconstructing past contexts. Conventional relevance‑based retrieval often overlooks three distinct aspects: semantic relevance, evidence‑grounded validity, and query‑specific eligibility. To address this, we introduce CALM (Conflict‑Aware Lifecycle Memory governance), a framework that decouples persistent memory preservation from query‑time admission, enabling conflict‑aware lifecycle management.\
CALM first converts heterogeneous records into evidence‑grounded memory cells and continuously updates each cell’s validity and lifecycle state using support, conflict, and dependency signals. During inference, the system admits query‑appropriate evidence via either a training‑free gate or a learned gate informed by temporal applicability.\
We evaluate CALM on the LongMemEval‑S and LoCoMo benchmarks using Qwen and GPT backbones. CALM consistently outperforms state‑of‑the‑art baselines in overall accuracy. Compared with simple filtering or reranking, the query‑conditioned admission further improves evidence selection. Conflict‑blocking and robustness analyses demonstrate that CALM suppresses erroneous retrieved memories while retaining useful evidence. Historical reconstruction experiments show that preserving superseded records supports accurate reasoning about earlier states without compromising answers for the current state.\
Review