Fluree launches an MCP-native memory layer for AI agents

Fluree launched Fluree AI on July 23, a serverless knowledge graph meant to give AI agents shared, persistent memory across sessions, per the announcement.

Right now most agent fleets run on separate contexts. No shared memory, no arbiter when they disagree. Ask two agents the same question and you can get two different answers, because neither one knows what the other queried or when. Fluree ties permissions to the data layer itself and exposes the graph as a first-class MCP tool, so any MCP client can query the same memory without needing a custom integration built for it.

The GA release ships with more than 300 connectors, including Salesforce, HubSpot, Snowflake, Postgres, BigQuery, Databricks and Stripe, per the announcement. That's the actual bet: an agent pulling customer data from a CRM and another pulling from a warehouse table should stop contradicting each other once they're reading from the same graph instead of two disconnected caches.

I've hit this same wall building agent workflows on my own. Session context dies the moment you close the terminal, and every new run starts from zero unless you've wired up your own memory layer by hand. A shared, permissioned graph that any MCP client can query without setup is solving a real problem, not a manufactured one. It's also the same siloed-agents problem HubSpot has been describing from the customer-data side, just attacked from the infrastructure side instead.

Fluree is betting that agents will trust a shared graph before they trust each other's output. But the announcement doesn't name a single customer running this in production, just a connector list and a serverless label. That's a spec sheet, not proof it holds up at real query volume.