Sparsh Jain pitched experiential memory as the 10x fix for repeat agent mistakes. Jain, chief of CentralAgent AI, made the case in an AI Engineer talk.
Before, the fix was to feed agents all context and memory. Jain said that approach raised cost, effort and time while lowering accuracy at enterprise scale. A senior executive at a large hedge fund told him course corrections vanished by the next session. Ten people could build two decades of combined experience with one agent, Jain said, and still watch it repeat day one errors.
His answer splits memory into short term working memory and long term memory. Long term memory divides into facts and experiential learnings, Jain said. Experiential learning further divides into case detail, strategy for a situation and generalized skill. As Jain put it, "people who do not remember from the mistakes or the agents who do not remember from their mistakes are condemned to repeat it."
Then came the lifecycle. Formation builds notes. Updation adds, revises or deactivates them. Retrieval decides performance. Jain warned against endless accumulation and cited a man with near perfect recall who grew overwhelmed by competing memories. Retrieval failures still kill good systems, he said, describing an agent that pulled maps for a New York home when the user wanted San Francisco. Context assembly fell from 20 minutes to 2 minutes per task after the fix, according to Jain.
Rest came next.
Jain described dreaming cycles that clean, prioritize and rearrange memory between sessions. Loops plus evals extract learnings after each session and test them before writeback, he said. Models in the past 6 months grew smart enough to act as verifiers, Jain said, removing humans from every review. Instead of writing prompts, teams now write loops, he said.
DeepSeek now proposes a separate coordination model where a central Router assigns work to parallel Workers, according to TechGig reporting. Jain closed with three rules for experiential memory, according to his talk. Disclose progressively, improve continuously and build loops, not prompts. He expects autonomous teams powered by shared agent experience to emerge in the next 6 to 9 months.
