Enterprise Brain – Memory that learns from work

Enterprise Brain is AmpUp’s system for giving AI agents shared, versioned memory grounded in a company’s existing context catalogs. It connects what an agent knew and did with the outcome, then turns graded experience into guidance for future runs. Teams can define the kinds of facts memory captures around their goals while keeping their existing catalog authoritative. Each task record is pinned to the memory version used, supporting evaluation and preparation of training data from the same history. Memory and trajectory records live in open Apache Iceberg tables, with Apache Arrow used in the processing engine; deployments can run locally or in a company’s AWS, Google Cloud, or Snowflake environment. Permissions, audit trails, snapshots, branches, and conflict-aware merges are intended to keep learning governed and under customer control.

AmpUp reports improved results across eight evaluations, with the largest gains where work or questions repeat: ORCA-bench incident root-cause success rose from 6.3% to 35.2%, while CRMArena-Pro accuracy increased from 57.6% to 76.7%. LongMemEval-S rose 21.3 points, and sales-pipeline question scores increased from 0.71 to 0.92. Other improvements were smaller: code review gained 3.2 points, SWE-ContextBench 3.0, and banking tasks 3.1; some tests also cost more. The evaluations compare against different baselines, and the page says the results measure learned memory and guidance, not models trained from the stored tables. It presents Enterprise Brain as a private-preview product.