[2/2] When One Analyst’s Memory Helps the Whole Team (Series #Agent)
Following up on the discussion of Databricks’ MemCurrent, I’d like to look at how memory can work for a user and at levels above the individual.
“Exclude background requests” is how an analyst corrects the active-user calculation in an example from the article. In that person’s next session, the test checks whether the agent uses the filter in a new report. When a colleague needs the same kind of report, how will the agent know to apply the filter for them too? By November 2026, analytics platforms already had several ways to share that knowledge with a team.
In Genie One, saved memories are personal. You configure shared context in Genie Agent. When you connect tables, it looks for popular queries on those tables among the ones the editor can access. The editor can accept suitable SQL as an example for future answers. A query already used by one analyst then becomes guidance for the shared agent. See the documentation on Genie One memory and setting up a Genie Agent.
Snowflake shows how the cycle works around the semantic layer. Cortex Analyst looks for recurring filters, metrics, and queries in the SQL history it can access. It also uses usage statistics to find popular questions without verified SQL. Its suggestions can be accepted into a shared semantic view. The tests have a nice detail. When a verified query is selected as the reference for an evaluation, Snowflake temporarily withholds it from the agent’s guidance and compares the result of the generated SQL with the reference.
Once the formula is agreed on, you can put it into the data model. In Databricks, one option is a metric view in Unity Catalog. In Microsoft Fabric, an agent can access a published Power BI semantic model and build DAX queries over its measures. The shared agent uses the model with the permissions of the person who asked the question. The agent’s own configuration, including instructions and source selection, can be versioned in Git.
Databricks also has a shared agent memory store on Lakebase for your own agent. The actor_id identifies whose memory it is: a user or another agent for personal memory, or a group ID for shared memory. You can optionally add a session_id to identify the session the memory came from. You can list, search, and write memories. Here you need to consider separately who can see the note itself, since it may already contain figures from the original query. The authors point out that anyone with access to the memory store can read all the memories. The actor_id has no effect on access rights.
In the MemAlign research project, Databricks tried extracting shared knowledge directly from conversations. The authors took logs from a working Genie Space, had an LLM judge select useful ones, and saved the knowledge they extracted. They report that after the first 62 entries, the test score exceeded 50%, compared with 33% for the version with expert instructions. This was an experiment in a single Genie Space.
Back to the example in the article, the background-request filter has a particular scope. In an activity report it helps count people. When investigating load, it may hide the very thing the analyst came to look for. When SQL from history becomes an example in Genie Agent or part of a Snowflake semantic view, the team has a specific place to see which data and questions the rule applies to.
#AI #Agents #Architecture #Data #Evals