memory that trains itself
edit: roadmap post. not a keynote. if you want the "we shipped it today" thread that's here.
Atulya started as memory infra. retain, recall, reflect, graph, Brain sniffing integrity stuff. fine when the agent is the product.
then every company hired "the fine-tune guy" and bought one A100 and a dream.
now you have two brains:
memory bank updates every time someone retains a ticket. model weights frozen at training_data_march_v7_REAL_FINAL.jsonl.
they drift. nobody owns the gap. ML owns the checkpoint. platform owns the bank. they meet quarterly and disagree politely.
someone asks in a review: "why does the model say eu-west when the agent literally has us-east in memory."
awkward silence. someone says "temperature." meeting ends. nothing changes.
Data Forge is the export half of fixing that. this post is the other half.