The use of electronic patients' record systems has been realized in many hospitals. Such systems make it easier to retrieve information than referencing paper records. Doctors and assistant medical staff must both refer to the same medical records for current patients. Medical records consist of sentences that have an attribute of clinical process. The clinical process cycle is comprised of three attributes: observation, diagnosis and treatment (hereafter abbreviated as “ODT”). Generally, the ODT cycle is a sequence of sentences that defines the structure of a record. The flow of attributes is interpreted as the context of a medical record. It is important for doctors to consider the context of each related document. However, if the records retrieved are not related to the context of the current matter, there is no need to refernce them. We propose a context—based similarity measurement model for retrieving medical records. The model exploits ODT cycles to calculate similarity values in the records. In a document, the model takes account of relations between sentences when two sentences have the same attribute (s) or sentences have a sequential attribute in a continuous ODT cycle. When two records have a homogeneous relation, the model measures the similarity between the documents. Then, similarity values between the records are referred to in a matching and ranking retrieval process. The proposed model was evaluated in two experiments. While there isn't much difference in matching between the model and a vector space model that measures similarity values by counting the term frequency included in two records, the proposed model is superior to the vector space model in terms of ranking. We conclude that the context—based model should be adapted in combination with the vector space model in order to execute effective retrieval.