Coworker is an AI teammate platform aimed at companies whose agents keep re-reading the same Slack threads, CRM records, docs, and Jira tickets on every prompt. Today's launch is OM2, an "organizational memory" layer that connects through a single MCP, ingests data from more than 50 enterprise tools, and distills it into a permission-aware graph of the customers, projects, people, and decisions that make up a business. The pitch to founders and operators is simple: over half of enterprise AI token spend goes into rebuilding context from scratch, and OM2 is designed to do that work once and reuse it. Under the hood, OM2 combines specialized language models and classical ML to extract discrete facts (who was involved, what was decided, which accounts are affected) and keeps refreshing them as new information lands in connected systems. It also converts repetitive tool-chain queries into deterministic lookups so common questions stop paying full LLM tax, and it pairs with Coworker's model routing to send simple retrieval to cheap models while reserving frontier models for real reasoning. Access controls are baked into every node in the graph, and the platform runs on single-tenant infrastructure with SOC 2, GDPR, and CASA Tier 2 posture. OM2 works with Claude, ChatGPT, Gemini, and open models hosted on Coworker itself, so teams are not locked into one vendor's memory. Coworker was founded by Alex Calder (CEO) and Bradford Church (CPO), both former Uber operators, and came out of stealth earlier this year with a $13M seed round to build a general-purpose AI teammate. OM2 is the follow-up to their original OM1 memory architecture, and it matters now because the industry conversation is shifting from bigger context windows to cheaper, governed retrieval. It is available today at coworker.ai.