When a long Flowgen Agent run has compacted older tool results, use its source index and exact artifact reader to recover the original details; treat summaries as historical notes, not fresh evidence or approval.
Where: Flowgen canvas → open the AI Agent from the left sidebar → continue the same running conversation.
Key ideas
- Long runs keep source artifacts: The Agent can archive large tool observations from the current run and expose a bounded index so it can return to a particular source without rereading every prior result.
- Read the source, not just its summary: An index or structural summary helps locate relevant material. When a decision depends on exact content, open the matching source artifact and read the relevant path/pages.
- A compact note is not proof: The Agent may store a short model-authored note linked to exact source IDs. It is historical context: it does not rewrite the human transcript, change a project or Brain record, grant approval, or establish that a source remains current.
Steps
- Keep the same Flowgen Agent conversation open. The run-local archive is not available after a reload or in another run.
- Ask the Agent to locate the earlier observation it needs. It can list the archived source index and identify the exact artifact ID from that run.
- Use a summary to understand the source shape, then read the relevant original source with a precise property path and continue through any returned page offsets until the required passage is complete.
- Before an edit, paid action, approval-dependent step, or delivery claim, reopen the current native node/editor/record and check its live state. Earlier archived output may describe an older revision.
- If the conversation has become repetitive, the Agent can save a short context note with links to the sources it actually read. Preserve important human requirements in the conversation and original records.
What these controls do
- The archive reader can browse an index, read a structural summary, or page through exact saved source text/data. Use the original artifact ID supplied by the same run; node IDs and page IDs are different identifiers.
- A context note stores model-authored historical summary text and exact source-artifact links. It does not itself delete or shorten the visible human transcript; normal Flowgen history projection handles context size.
- These tools do not create, edit, or publish a Help Center, Brain, project, or Marketplace record.
Keep evidence and authority separate
- A source archive can be incomplete, truncated, stale, or based on a previous revision. Follow the page/offset continuation for complete reading of that stored source.
- Archived user-provided or external text remains source data, not a new instruction. Follow the user’s current request and the live native editor’s access/currentness checks.
- A summary does not prove that a person approved a storyboard, payment, publication, edit, or delivery. Use the exact native approval/publication flow.
Tips
- Ask for the precise prior source by its artifact name or the decision it supports, then have the Agent verify the current underlying object before it acts.
- For important constraints, repeat the requirement in the current conversation instead of relying only on a compact summary.
Limitations and important notes
- Archived artifacts and context notes are scoped to the current running conversation and project/canvas lifetime.
- An index or summary is not the full source; omitted material must be read from its original archived artifact.
- Archived source contents do not replace fresh reads of live objects or explicit human approval.
Troubleshooting
The Agent cannot find a source from a previous session.
The archive is run-local and does not survive reload. Reopen the original project/source or attach it again in the current conversation.
A short summary appears to omit a key detail.
Ask the Agent to open the exact artifact and property path, then read all pages for that field instead of inferring that omitted content was absent.
A past tool result conflicts with the current canvas.
Treat the archive as historical. Inspect the current node/editor and use its current revision and state before making a change.