As large language model (LLM) agents become more active in persistent social environments, assessing their long‑term interaction abilities is increasingly critical. Existing work has largely advanced long‑term memory and conversational context retrieval, but has paid little attention to whether agents can autonomously re‑engage with conversations that have become inactive. Remembering a past interaction does not guarantee that an agent can recognize when to revisit it, decide to return, or rebuild the context needed for a coherent continuation. To address this gap, we define agent social re‑entry as a multi‑dimensional evaluation problem, separating three complementary capabilities: Finding (identifying the target conversation), Returning (initiating a target‑directed reply), and Reconstructing (restoring the conversational context). We introduce a shared‑snapshot evaluation framework that measures each dimension independently, allowing us to study how conversational resurfacing and persistent memory relate to different aspects of social re‑entry. Experiments across six LLM configurations show that resurfacing markedly improves Finding, while persistent memory correlates with higher rates of autonomous target‑directed replying (Returning). However, neither improvement yields consistent gains in Reconstructing. These findings highlight the distinction between accessing prior conversations, choosing to re‑engage, and maintaining conversational continuity. Our framework enables separate evaluation of these capabilities without relying solely on memory retrieval or response quality metrics.
Review