Finding
Long Hermes research and coding runs become fragile when context management depends on the model surviving a huge transcript instead of deliberately preserving decisions, artifacts, and reusable procedure outside the live context window.
Current
A real Hermes installation can handle complex multi-step work with large context, compaction, session history, files, skills, memory, and delegation. The weak point appears when long investigations keep accumulating raw tool output, repeated explanations, partial decisions, and temporary notes in the chat itself. Without a context handoff habit, later turns may lose the “why,” repeat discovery work, or compress away the exact verification steps needed to continue safely.
Suggested
- Add a long-run context checkpoint rule. Exact change: add a “Context Dragon checkpoint” section to
SOUL.mdor the main operator runbook: after long research, coding, debugging, or multi-agent work, write a compact handoff containing goal, current decision, key evidence, changed files or artifacts, open risks, and next verification step before continuing. - Move large working knowledge into artifacts instead of chat. Exact change: update the research/build runbook with: “If notes, source maps, implementation plans, or test findings exceed a short summary, save them as a file artifact or wiki page and keep only the path plus a 3-line summary in the conversation.”
- Convert repeated context into skills or memory boundaries. Exact change: patch the Optimizer Agent review prompt with: “When a session shows repeated context overflow, compaction pressure, or rediscovery of the same procedure, recommend either a skill patch for the workflow, a compact memory for stable facts, or an artifact for large reference material.”
Impact
This makes long Hermes work restartable instead of dependent on one uninterrupted context window. It reduces token waste because future turns can load the right skill, artifact, or session summary instead of re-reading the full transcript. It also improves reliability during compaction: the important operational state is preserved in a deliberate handoff rather than buried in raw chat history.
Effort
Small — the main change is adding one checkpoint habit, one artifact-saving rule, and one optimizer review trigger. No new infrastructure is required because Hermes already has skills, memory, files, session history, and compaction-aware workflows.
Public page note
Safe public content includes the context-management principle, generic handoff checklist, artifact-vs-memory-vs-skill boundary, and maturity benefits of shorter live context. Internal-only content includes raw transcripts, private research notes, exact local paths containing sensitive material, credentials, logs, environment values, customer data, and unfinished implementation details.