In everyday social situations, people use knowledge about knowledge of their interaction partners (partner models). For example, students need to encode and remember whether their learning partner knows a lot or only a little about a specific learning topic. Partner models help to make informed judgments about whether learning partners can provide help or need explanations. In a preregistered experimental study, we examined how partner models about a learning partner are formed in a simulated computer-supported collaborative learning context. We applied multinomial processing tree models to measure memory processes and guessing biases separately. N = 168 participants were first assigned to a simulated learning partner (partner schema: expert, intermediate, or novice). They then received information on that partner’s specific knowledge levels (high, medium, and low) for learning topics, before retrieving their learning partner’s knowledge level for each topic. For the expert and intermediate partners, high and low levels were better remembered than medium levels. However, for the novice partner, only low levels were better remembered than medium levels. Models for experts and novices were more accurate than for intermediates. Generally, metacognitive judgments were not always in line with actual memory. When topics and levels were not remembered, guessing processes showed a schema-consistent bias: experts were overestimated and novices were underestimated. Memory and guessing processes involved in partner modeling are shaped by social schemas. Beyond informing the design of educational tools, the findings can generalize to contexts in which judgments about others’ knowledge guide everyday social decision making.