The Datomic database uses a memory store just for easiness of setup. After all, this is just for demonstrational purposes.
The io.rkn/conformity library has been used to ensure idempotency of operations when transacting the Datomic schema and sample data.
No error handling has been implemented for lack of time. A good choice for implementing validation and error reporting is clojure.spec and maybe phraser for producing nice error messages for the end user.
Each Lacinia resolver makes a query to the Datomic db. This is an ok solution when the Lacinia server is also a Datomic peer. Not that good when querying Datomic as a datomic client.
The Big Kahuna is translating Graphql queries to Datomic pull queries. That would allow to fetch all attributes in one go with one single pull request instead of quering the database whenever an attribute needs to be resolved. This is by the way mitigated by Datomic because of the good data locality it exposes for related datoms. The Datomic's peer caching strategy should kick in and avoid network round trips in many cases.
Anyway, the "Graphql->Datomic pull request" translation layer would allow to traverse the entity graph without making multiple (and in many cases costly) queries on big databases.