Blog
Technical deep dives, product updates, and practical engineering insights for Data + AI teams
MatrixOrigin Hosts AI Leadership Dinner in Silicon Valley: From Compute to Intelligence to Productivity
MatrixOrigin hosted an AI Leadership Dinner in Silicon Valley, welcoming a delegation of Chinese AI industry companies visiting NVIDIA headquarters and bringing industry partners together to explore enterprise AI adoption and collaboration.
MatrixOrigin Hosts AI Leadership Dinner in Silicon Valley: From Compute to Intelligence to Productivity
Agents Learn. Bad Memory Scales Too.
Agents Learn. Bad Memory Scales Too.
Claude Code Projects and procedural memory are pushing agent experience closer to execution. This article examines why shared, writable memory needs versioning, provenance, conflict handling, and rollback.
The Agent Found the Right Answer. JEV Rejected It.
The Agent Found the Right Answer. JEV Rejected It.
This article uses an Agent Memory experiment to examine the accuracy, cost, and failure modes of a JEV judgment layer, and explores when adding a helper model to an Agent is actually worthwhile.
From App + Database to Agent + Context: Rethinking the Architecture of Enterprise AI Infrastructure
From App + Database to Agent + Context: Rethinking the Architecture of Enterprise AI Infrastructure
Enterprise AI must connect two worlds: deterministic software that executes reliably and probabilistic agents that reason, use context and learn from feedback.
From China to the World: MatrixOrigin Joins the 2026 Google for Startups Accelerator China
From China to the World: MatrixOrigin Joins the 2026 Google for Startups Accelerator China
MatrixOrigin has been selected for the 2026 Google for Startups Accelerator China as one of 10 companies chosen from several hundred applicants.
Astra Is Now Open Source: An Enterprise Agent Runtime for Long-Horizon Work
Astra Is Now Open Source: An Enterprise Agent Runtime for Long-Horizon Work
MatrixOrigin is open-sourcing Astra, a self-hosted, model-vendor-neutral Agent Runtime for enterprise work: long-horizon tasks that advance on fewer tokens, agent changes that can be traced and rolled back, and one piece of work that follows the user across environments.
AI-Native Organization in Practice: From FDE to FDB to an OPC Incubator
AI-Native Organization in Practice: From FDE to FDB to an OPC Incubator
MatrixOrigin shares how it moved from FDE to Forward Deployed Builder (FDB), and why an AI-native organization should become a virtual incubator for builders, projects, and new businesses.
MatrixOrigin Recognized in the 2026 Gartner® Coolest Vendor Innovations in Data Management
MatrixOrigin Recognized in the 2026 Gartner® Coolest Vendor Innovations in Data Management
MatrixOrigin has been recognized in the 2026 Gartner® Coolest Vendor Innovations in Data Management report as the only China-based vendor among the four companies named. MatrixOne Intelligence unifies databases, AI workflow orchestration, agent runtime observability, and persistent memory to provide a reliable, cost-effective foundation for enterprise AI applications.
Good News | MatrixOrigin Named a 2026 Shenzhen "Potential Unicorn Enterprise"
Good News | MatrixOrigin Named a 2026 Shenzhen "Potential Unicorn Enterprise"
MatrixOrigin has been named a 2026 Shenzhen Potential Unicorn Enterprise, recognizing its technical innovation, growth potential, and market value in Data & AI.
MatrixOrigin Raises $10M Series A to Scale Enterprise AI Infrastructure
MatrixOrigin Raises $10M Series A to Scale Enterprise AI Infrastructure
MatrixOrigin has completed a Series A funding round of more than $10 million, backed by a HAND-led fund, AsiaCom, and Artesian Venture Partners. The funding will accelerate MatrixOne Intelligence development, global expansion, and enterprise AI deployment at scale.
MatrixOne Git4Data Deep Dive (Part 13) · Agents — Memory: From the Industry's Approaches to Governable Long-Term Memory
MatrixOne Git4Data Deep Dive (Part 13) · Agents — Memory: From the Industry's Approaches to Governable Long-Term Memory
Agent memory isn't a longer context window — it's the state layer that lets an agent keep working across tasks. This piece starts from what memory is, surveys the industry's approaches (prompt files, summarisation, vector retrieval, structured stores, platform built-ins), then shows how MatrixOne combines structured data, hybrid retrieval, and the branching, DIFF, snapshots and rollback of its Git4Data capability into long-term memory that is retrievable, auditable and recoverable.
MatrixOne Git4Data Deep Dive (Part 12) · Large Models — RLHF Preference Data: Disagreement, Adjudication, Reproducibility
MatrixOne Git4Data Deep Dive (Part 12) · Large Models — RLHF Preference Data: Disagreement, Adjudication, Reproducibility
Git4Data Part 12: preference data's unit is a pair, not a row, and it is computed from annotator votes rather than collected — the same data feeding both the RLHF and DPO routes. From 63,000 votes this derives preference pairs, audits degenerate pairs, no-consensus, preference cycles and length bias on a branch, materialises the conflict list before adjudicating on branches, and binds the dataset to its reward model. Verified on MatrixOne 4.1.0.
MatrixOne Git4Data Deep Dive (Part 11) · Large Models — SFT Data Curation: Auditable and Reproducible
MatrixOne Git4Data Deep Dive (Part 11) · Large Models — SFT Data Curation: Auditable and Reproducible
Git4Data Part 11: SFT data is orders of magnitude smaller than pretraining data, so every curation decision imprints on model behavior. Using one chat model's SFT pool, this runs a full curation pass on a zero-copy branch — exact dedup, near-dup, quality gate, safety, benchmark decontamination, multi-turn integrity — counting before each filter, with DATA BRANCH DIFF reporting the net change, then register-swap-snapshot to release. SQL verified on MatrixOne 4.1.0.