Databricks Says Memory Triples an Agent Pass Rate: MemCurrent Lifts 22.3 to 68.4 Percent on Its Own Benchmark

Databricks has published the design of MemCurrent, the memory system now running behind its Agent Bricks Memory product and its Genie data assistant, together with results it says show agent memory is worth more than most teams expect. The premise is familiar to anyone who has watched a coding agent start over: an agent forgets everything when a session ends, then relearns your table names, your corrections and your preferences on the next run. Memory turns that into a notebook the agent writes while it works and reads before it starts.

The design splits memory into two kinds of notes. One is a raw episodic 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 end, extracts lessons, merges duplicates and resolves contradictions — the expensive judgment work of deciding what to keep, which the company names as the hard part of agent memory.

The headline numbers come from an internal benchmark Databricks built from real Genie conversations: 114 test cases, 29 users and 2,233 past sessions. The company reports that MemCurrent raised the pass rate from 22.3 percent with no memory to 68.4 percent with it — roughly triple — while making runs 1.35x faster and 1.44x cheaper, at a median 252 seconds and 418K tokens. The best rival it tested, the open-source memory system Mem0, reached 43.9 percent, and a ceiling case that injects only the gold-standard memories scored 78.6 percent. All of these figures are Databricks self-reported, with no independent replication cited.

On three public memory benchmarks — LoCoMo, LongMemEval and ConvoMem — the company reports MemCurrent averaging 88.1 percent accuracy, marginally above the 87.8 percent achieved by stuffing every past session into the prompt, while reading about 8.4K input tokens per question instead of 103K, roughly twelve times fewer. On the two benchmarks with the longest histories, MemCurrent beats full context outright, because a prompt crammed with hundreds of stale sessions confuses the agent. Both agent and memory model were GPT-5.6 Luna, evaluated across LangGraph and OpenAI Agents SDK harnesses with three repeats.

The ablations are where the paper gets interesting for anyone building memory. Distilled notes alone lifted pass rate to 66.1 percent, while raw session logs alone reached only 35.8 percent — distillation, not volume, does the work. The background dreaming pipeline alone scored 63.5 percent against 53.9 percent for memory written on the spot through tools. Databricks also offers a concrete anecdote: a first attempt at a "top five SDP customers" question took eleven tool calls, eight subagents and nearly six minutes; the repeat, drawing on saved lessons, needed two tool calls, one subagent and 41 seconds.

The commercial argument rides on portability. Because memory is stored as plain files rather than hidden inside a model, Databricks says it can follow a user to a different agent or a different LLM, and potentially be shared across an organization — a knowledge asset that outlives any single model version. That framing matters in a market where OpenAI, Anthropic and a cluster of startups including Mem0 and Letta are all converging on persistent agent memory as a retention layer.

The caveats are the usual ones for vendor research: the internal benchmark is small, drawn from one kind of work — people asking questions about their company data — and every number comes from the company that sells the system. The honest test for any team paying for long agent runs is to replay a few hundred of its own sessions with and without memory. But the directional claim, that what an agent remembers matters as much as which model it runs on, is becoming the consensus position of 2026 — and the race is now over who gets to hold the notes.