Docs / Getting started / Threads & memory

Threads & memory

CloudyBot organizes work in threads — one conversation per job or topic in the dashboard. Memory is per thread, not global across your account: a pricing research thread does not automatically know what you said in a hiring thread. Your ~/files/ workspace is shared; conversation context is not.

Prompt tips: sending messages · first run: your first task.

What a thread remembers

• Your messages and assistant replies in that thread, including tool traces (browser steps, file paths, integration reads).
• File references — filenames and outputs from earlier turns, so follow-ups like “add a column to that table” work without re-uploading.
• Saved facts you explicitly ask CloudyBot to remember for this thread (templates, standing preferences, glossary terms). These are injected on every turn until you remove them.

To save a fact, say it plainly: “Remember: our Q3 launch date is 15 September — use that in future replies in this thread.” For standing rules that belong on a role, put them in Specialist instructions instead (see below).

Rolling context (long threads)

Very long threads are summarized so the model stays within context limits. Recent turns stay detailed; older back-and-forth collapses into a short summary. The dashboard UI still shows your full transcript — summarization affects what the model sees on the next message, not what you can scroll back to.

Practical effect: after many turns, the agent may forget exact wording from early messages but should retain the gist. If a detail matters later, either repeat it, save it as a fact, or start a fresh thread with a one-line recap at the top.

When to start a new thread

• New topic — unrelated job (don’t mix “weekly SEO report” and “invoice cleanup” in one thread).
• Clean slate — the agent keeps misreading an old assumption; a new thread avoids carrying bad context.
• Reusable template — you perfected a prompt and want a fresh run without prior tool noise.

Stay in the same thread when iterating on one deliverable — filenames, URLs, and table structure from earlier turns are already loaded.

Specialist threads

When you chat with a Specialist, each thread also loads that role’s instructions and enabled skills. Scout, Watchdog, and similar presets bring their own playbook; custom Specialists use duties you define. If behavior drifts, tighten the Specialist settings or open a new thread tied to that Specialist.

Specialist memory does not leak across Specialists — only the thread’s own history and saved facts apply.

Scheduled tasks and cron

Scheduled tasks run against a target thread (or create context on each run, depending on how you configured the duty). Cron output appends to that thread’s history, so the next manual message can reference yesterday’s automated report. Keep recurring jobs on a dedicated thread per duty so summaries stay focused.

What threads do not share

• Conversation history or saved facts from other threads.
• Browser session state after idle timeout — long gaps may start a fresh browser session on the next tool use (your files in ~/files/ remain).
• Billing quotas — AI credits and browser minutes are per account, not per thread (limits FAQ).

Privacy and retention

Thread data is tied to your account. Deleting a thread removes its conversation row; saved facts for that thread go with it. Account-level policies: data & privacy FAQ · Privacy Policy.

Developers: Memory model (technical overview) · Scheduled tasks · Troubleshooting