All types of databases, Redis, Postgres, Clickhouse, Mysql etc. now display important stats and metrics aswell as most run queries in the Maple service map
Google's onto something!
They just made every other approach to agent memory look incomplete.
When a stateless agent crashes mid-workflow and restarts, it loses all its progress and decisions.
And because LLMs are non-deterministic, the restarted agent will likely make different intermediate choices on the same data.
Imagine an agent who processed financial records over five days.
- On day 1, it ingested 4,000 records and normalized formats. One record has an ambiguous date field ("03/04/2026"). Say the agent interpreted it as March 4th. Every downstream decision builds on that interpretation.
- On day 3, the agent crashed.
A stateless restart means the agent re-ingests all 4,000 records from scratch. This time, the LLM could interpret that same ambiguous date as April 3rd.
The final output will be different, not because the data changed, but because the reasoning chain diverged.
This is fundamentally different from a database crash. Database replays are deterministic.
To actually solve this problem, the agent needs to periodically save its intermediate state, including the progress, accumulated decisions, and the reasoning chain that led to them.
If it crashes, it reloads the last checkpoint and continues from there with every prior decision intact.
If you want to see this in practice, Google Cloud's Agent Platform implements this through three mechanisms, and they have partnered with me today to show what they are doing!
→ Memory Bank gives agents a persistent state that accumulates across sessions over days or weeks. The agent on day 5 can easily read what was decided on day 1 without re-processing anything.
→ Resume Agents handle checkpoint-and-resume natively. When an agent crashes or pauses for human approval, Agent Runtime saves the full state. When it resumes, the context reloads exactly as it was. The agent consumes zero compute while it waits.
→ Ambient Agents make execution event-driven. The agent activates when new data arrives, not when a human types a prompt.
The hardest problem in production agents isn't making them smarter. It's making sure they remember what they already decided.
I have shared the link to the Agent Platform in the replies!
We made a collection @GoogleDeepMind scientific agent skils for research tasks, genomics, structural biology, cheminformatics, literature search, and more.
👉https://t.co/zkPuCtmwEE
https://t.co/zkPuCtmwEE