Next Wave for Agents: Runtime Environments Plus Memory Systems

青稞AI · wechat · 2026-07-15

This article argues that the next notable tech wave in the Agent space isn't isolated tool capabilities, but a combination of distributed runtime environments + distributed memory systems.

Why These Two Directions

Runtime Environments solve where Agents execute: local/cloud, isolated, elastic scaling, and long-running tasks.
Memory Systems solve what Agents remember: context continuity across sandboxes, sessions, and multi-Agent collaborations.
The author emphasizes that while many systems can now run tasks for a long time, memory remains stuck in shallow solutions like "vector DBs / Markdown memos," failing to keep up with increasing task durations.

Key Observations from the Article

E2B data shows that from 2024 to 2025, the average runtime per sandbox grew by over 10x, indicating Agents are shifting from short to long tasks.
Practices and products mentioned include Manus, domestic cloud sandbox solutions, and Tencent Cloud's open-source CubeSandbox.
The author categorizes memory systems into three layers: persistent, shared, and evolving. The "evolving" layer means memory shouldn't just archive but also extract patterns and preferences from historical tasks.

Conclusion

The article compares Agent platforms to early cloud computing: we now have "EC2-style" runtime environments, but lack an "EBS-style" persistent memory layer for Agents. The author predicts this gap will soon become the next major focus for infrastructure innovation.

Original post →

Next Wave for Agents: Runtime Environments Plus Memory Systems

More from coding & agent