| name | context-canvas-memory |
| description | Use when a Hermes Agent task becomes long, tool-heavy, evidence-heavy, or likely to lose context. Maintain a compact Task Canvas with raw evidence refs instead of carrying every log in active context. |
| version | 1.0.0 |
| author | hermes-agent-harness-plus contributors |
| license | MIT |
| metadata | {"hermes":{"tags":["context-management","task-canvas","evidence","mcp","long-running-tasks"],"related_skills":[]}} |
Context Canvas Memory
Overview
Use a Task Canvas as a short-term working map for long Hermes Agent tasks. The
canvas keeps concise nodes in JSON, while raw logs, diffs, and command outputs
live in referenced evidence files.
The goal is not to remember everything forever. The goal is to keep the agent
from losing the shape of the work while preserving a path back to the facts.
When to Use
- A task has five or more tool calls.
- Logs, diffs, browser output, or research notes are too large for active
context.
- The agent has tried several paths and needs to avoid repeating dead ends.
- The final answer must cite what was actually checked.
Do not use it for secrets, private memory dumps, or simple one-shot questions.
Basic Workflow
- Start a canvas with a one-sentence goal.
- Add raw evidence refs for important tool outputs.
- Add or update short nodes that point to those refs.
- Mark failed paths as blocked or deprecated instead of leaving stale notes.
- Close out with decisions, verification, and follow-up items.
Invariant
Every factual finished node should follow this chain:
node summary -> evidence ref -> original/verifiable content
If a note has no evidence, mark it as a plan, question, or assumption.
Common Pitfalls
- Treating Mermaid as the database. Mermaid is a display projection; JSON is the
source of truth.
- Writing confident summaries without evidence refs.
- Capturing everything. A canvas is a map, not a transcript.
- Saving temporary task scratch into durable memory.
Verification Checklist
