How to start a new conversation with OpenClaw — and what it remembers

/new archives and starts fresh, /reset resets in place, and workspace memory survives both. Plus 2.0's quiet change: the daily 4 a.m. auto-reset is gone, so sessions now run until you end them.

The short answer

Type /new to archive the current conversation and start a fresh one, or /reset to reset the one you are in, in place. Sent alone, either is acknowledged without waking the model. What survives the reset is the part most owners actually worry about: the agent’s workspace memory (MEMORY.md, the memory/*.md files, and anything it wrote to disk) persists across every reset; what goes is the transcript’s presence in active context. And since OpenClaw 2.0, nothing resets on its own: the old 4 a.m. daily reset is no longer the default, so a conversation now runs until you end it or configure an expiry yourself.

/new and /reset, precisely

/new archives the current session and starts a fresh one. /new <model> starts the fresh session already switched to that model: an alias, provider/model, or a fuzzy provider name all resolve; text that matches no model is treated as an ordinary message. In the Control UI, a typed /new creates and switches to a fresh dashboard session, except when session.dmScope: "main" is configured and you are in the agent’s main session, where it resets that session in place instead. A direct message rides the main session by default, and on a Telegram chat with one of our own 2026.9.1 pods in September 2026 that is what /new did: a fresh context in the same session, with a model picked by /model -s carried over. The way back to the default is /model default -s.

/reset resets the current session where it stands. /reset soft is the gentler variant: it keeps the transcript, drops reused CLI-backend session ids, and reruns startup. That is useful when a backend wedged but you do not want to lose the conversation record. Both commands work for channel-authorized senders, not only the owner, on channels that do not enforce owner-only commands.

What the agent actually remembers afterwards

A reset replaces the session id and stops loading that transcript into context. It does not touch the agent’s memory. Shared workspace memory (MEMORY.md and the memory/*.md files) keeps its existing contents, and anything the agent saved through a tool (notes it wrote, files it created, a decision it recorded because you asked) is ordinary disk and survives untouched. So the practical rule: facts the agent stored carry forward; the running thread of the old conversation does not, except as far as memory search reaches it. With rememberAcrossConversations enabled under memory.search, the agent can also recall material from your earlier private conversations with it. That is recall, not context: retrieved when relevant rather than always present. If something must survive a reset verbatim, ask the agent to save it before you type /new; how the memory layers fit together is its own guide.

Nothing resets on its own anymore

Through the 2026.7.x line the default was a fresh session every day at 04:00 local. OpenClaw 2.0 changed the default to none: sessions persist until you reset them. Explicit policies still work, and ones you had configured are kept across the upgrade:

"session": {
  "reset": { "mode": "daily", "atHour": 4 }   // or:
  "reset": { "mode": "idle", "idleMinutes": 120 }
}

With both configured, whichever expires first wins. If you upgraded and your agent suddenly seems to remember yesterday’s conversation where it used to start clean, this default change is what you are noticing: it is one of 2.0’s quieter behavior changes.

Where sessions live, and what archiving means

Since 2.0, session rows and transcripts live in SQLite, with archived transcripts beside the store:

~/.openclaw/agents/<agentId>/agent/openclaw-agent.sqlite   # live sessions
~/.openclaw/agents/<agentId>/sessions/                     # archives

Archiving is not deletion: dashboard sessions auto-archive after seven days of inactivity by default, other session types archive only when you say so, and archived sessions are user-protected, exempt from every automatic maintenance path. The 2.0 Control UI can also search past transcripts by exact words, which changes what “losing” a conversation to a reset costs you: the record stays findable even once it is out of context.

The cost of never resetting

A session that never ends keeps growing, and context is what you pay for: the longer the running conversation, the more every turn carries. Under the no-reset default, compaction manages the active context as it grows, summarizing older stretches, and /compact triggers it on demand. /status shows current context usage when you want the number. A reasonable operator rhythm on 2.0: let working conversations run, save anything durable into memory, and start fresh with /new when a thread genuinely changes subject: the model thinks more clearly in a context that is about one thing, and your per-turn cost resets with it.