Hannes Rudolph and Patrick Erichsen interview Peter Steinberger about the architecture behind a forthcoming OpenClaw release. He describes moving session state from scattered JSONL access toward a shared SQLite interface to support longer tasks and cloud-backed work. The discussion records development plans at the time of the episode, not a claim that every preview feature is now available.
Peter Steinberger describes a team server where trusted maintainers can see and continue shared sessions. The aim is to avoid repeatedly copying agent responses between people and tools. He distinguishes internal team sharing from sending arbitrary outsiders a session link, and discusses the challenge of preserving context without leaking information across conversational boundaries.
Peter Steinberger discusses more opinionated defaults, pluggable memory and background memory processing. The group acknowledges a skill-maintenance incident and the risk that incorrect persistent skills can poison later work. He favors separating a planning session from worker sessions and choosing models according to the task; reported cost and speed improvements are anecdotal rather than a general benchmark.
Peter Steinberger outlines easier onboarding, permission-aware discovery of existing tools, browser-control options and experimental node hardware. Claims that prompt injection is largely solved are not established by the conversation and should not be taken as a safety guarantee. The speakers also note migration edge cases, unfinished experiments and the need for human architectural judgment; brand disputes, merchandise promotion and personal banter are omitted.
Watch on YouTube
