Databricks introduces AI memory that triples task pass rates in its tests

AI agents can carry a correction into later work without rereading every past conversation. Databricks AI Research introduced MemCurrent on October 20, reporting that it more than tripled task pass rates while reducing token use and wait times in an internal test. The persistent-memory architecture now powers Agent Bricks Memory and Genie, the company says.

This is memory outside the model, not retraining. In its earlier memory-scaling research, Databricks described a persistent store that agents consult while working, separate from model weights and the current conversation.

Keep the record, revise the lesson

MemCurrent keeps two kinds of memory: raw session records and distilled lessons in a structured knowledge base. Both live in a database. The lesson store behaves like editable files, with an automatically generated index describing their contents.

The agent receives that index at session start, then can search both memory types and fetch complete entries. Tools let it add, update or delete memories following explicit instructions.

An offline pipeline also extracts implicit lessons from new sessions without an explicit request to save them. It splits oversized files, merges duplicates and resolves conflicts. Tool updates take effect immediately; background updates arrive with a delay.

Testing whether recall improves the work

Databricks built its Internal Memory Benchmark from real Genie conversations. Automated checks and manual review produced 114 test cases covering 29 users and 2,233 historical sessions. Tasks required applying past knowledge, not merely retrieving a phrase. A language-model judge scored the outputs.

MemCurrent averaged 418,000 tokens and 252 seconds per task—31% fewer tokens and 26% faster than no memory. These are company-run measurements inside its internal Genie agent, which generated the benchmark conversations. Supplying only the correct memories directly produced a higher pass rate of 78.6%.

On LoCoMo, LongMemEval and ConvoMem, which test recall from synthetic conversations, Databricks reports 88.1% average accuracy versus 87.8% when supplying every past session. Average input use was 8,400 tokens per question versus 103,000. Those tests used two agent frameworks and three repeats.

Remembering how an answer went wrong

One internal example shows the intended benefit. An agent wrote a correct database query but invented the answer table without reading the saved results. After user feedback, it stored the lesson. On the next attempt, Databricks says, it read the results before answering correctly.