Databricks says an AI agent that remembers its past work passes three times as many tasks as one that starts blank every session. The company’s research team described MemCurrent, the memory system now running behind Agent Bricks Memory and its Genie data assistant, in a blog post this week.
“Agent memory” is easy to explain. Normally an agent forgets everything when a session ends, so it has to relearn your table names, your corrections, and the way you like results presented. Memory gives it a notebook: it writes lessons down while it works and reads them back the next time. The hard part, Databricks writes, is deciding what to keep, because piling in every past conversation costs money and can bury the agent in stale or conflicting notes.
MemCurrent keeps two kinds of notes. One is a raw log of each session. The other is a set of distilled lessons stored as files, with an index the agent reads at the start so it knows what it already knows. Explicit instructions go in through tools the agent calls itself. Everything else is handled by what Databricks calls a “dreaming” pipeline, which runs in the background after sessions, pulls out lessons, merges duplicates, and resolves contradictions.
The headline numbers come from an internal benchmark Databricks built from real Genie conversations: 114 test cases, 29 users, and 2,233 past sessions. Databricks reports that MemCurrent raised the pass rate from 22.3 percent with no memory to 68.4 percent. It also reports 31 percent fewer tokens and a run that is 26 percent faster, which the post restates as 1.35 times faster and 1.44 times cheaper. The best rival it tested, Mem0, reached 43.9 percent. An ideal setup that hands the agent only the relevant memories scored 78.6 percent.
On three public memory benchmarks, Databricks reports that MemCurrent averaged 88.1 percent, level with the 87.8 percent from stuffing every past session into the prompt, while reading about 8.4K input tokens per question instead of 103K.
Databricks ran these evaluations itself, on its own product and a test set it designed, and the post cites no independent replication. The benchmark is also small, and drawn from one kind of work: people asking questions about their company data.
The post’s best evidence is anecdotal but concrete. In one Genie session, a user asked for the top five “SDP” customers. The agent had to work out what the acronym meant, find the right tables, and build the joins. That ran to eleven tool calls, eight subagents, and almost six minutes. It saved the steps as a workflow. When the user later asked a related question, the repeat run needed just two tool calls and a single subagent, and finished in 41 seconds.
The ablation tests show where the gain comes from. Distilled notes alone lifted the pass rate to 66.1 percent, while the raw session log alone reached only 35.8 percent. Background dreaming (63.5 percent) beat on-the-spot tool writes (53.9 percent) by nearly 10 points, though combining the two did best.
Databricks also argues that memory kept as plain files can follow a user to a different agent or a different model. That matters commercially: if the notes outlive the model, the vendor holding them gains a reason to keep customers.
A team paying for long agent runs should replay a few hundred of its own sessions with and without memory before trusting a 3x claim from a vendor’s own test.
Reported by Databricks in its blog post “MemCurrent: Designing Persistent Memory for Production Agents”, published on 20 October 2026.
