Agent Memory Is Solved. Agent Learning Isn't. — Karthik Ranganathan, Yugabyte
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
Karthik Ranganathan, co-founder and co-CEO of database company Yugabyte, argues in an AI Engineer talk that one remembering is largely solved, but a team of agents learning together is not. His stated problem is handoff: when one agent passes work to another, it sends the output but loses the reasoning and dead ends behind it, so the next agent spends tokens re-deriving them.
Yugabyte's answer is Meko, its own product, which stores that is governed, promotable from private to shared use, and traceable to its source. The description also reports that fixing the company's retrieval pipeline one variable at a time took faithfulness (how well answers stay in retrieved material) from 14% to 82% with about one-seventh the context.
Those results are the speaker's and the vendor's, taken from the talk's description rather than a published evaluation. The talk is best read as a statement of the handoff problem and one company's approach.
- Ranganathan argues that more memory, shared files or bigger context windows do not add up to learning; shared context needs to be governed, promotable and traceable to its source.
- Yugabyte says it fixed its own pipeline by changing one variable at a time, raising faithfulness from 14% to 82% while using about one-seventh of the context.
- The demo shows Meko, Yugabyte's agent-native persistence layer, resuming sessions across agents such as Claude and Codex and promoting private memories into shared data packs; the talk treats human promotion as a training signal.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
- RAG — Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.
Checking sign-in…
Loading comments…






