Spatial reasoning relies not only on metric attributes such as distance, angle, and shape, but also on topological relations that stay invariant under continuous deformation. Cognitive science treats these relations as fundamental to spatial understanding, yet most evaluations of foundation models focus on metric or viewpoint‑dependent relations.\ \ We introduce MindTopo, a benchmark that tests topological intuition across five properties grounded in cognitive science and formal topology: continuity, separation, order, enclosure, and knots. Each property is examined at two cognitive levels. Reasoning tasks ask a model to identify topological relations or infer how they change; planning tasks instantiate the model as a closed‑loop agent whose policy selects actions in an environment.\ \ MindTopo comprises 11,030 instances spanning 13 procedurally generated task types with controllable difficulty. We benchmark 14 multimodal large language models (MLLMs) and study agent configurations augmented with image and video generation, including three video generative models in planning settings. Results show that every MLLM performs better on reasoning than on planning, and the best model remains far below human performance. Supervised fine‑tuning and reinforcement learning improve reasoning more than planning for Qwen3‑VL‑2B‑Instruct. Generated observations retain local cues and reach plausible endpoints, but audited rollouts do not reliably follow environment dynamics nor preserve topology across transitions.\ \ Visit https://mind-topo.github.io/ for details.\ \ Review: Current foundation models still struggle with topological reasoning, especially when continuous action planning is required. Maintaining global topological constraints across state transitions remains a challenge, indicating a need for explicit modeling of topological invariants in future work.