Context Editing Memory Tool

Instructions

Use this skill to keep long-running agents effective as tool results accumulate.

1. Classify new information as one of:
   • Ephemeral: needed only for the next 1–3 steps.
   • Session state: needed until the current workflow completes.
   • Long-lived knowledge: needed across sessions (decisions, constraints, learned facts).
2. Keep context minimal: leave only what is required to reason about the next steps.
3. Persist long-lived knowledge to memory (file-based storage you control).
4. Write memory entries like documentation:
   • Title
   • Date/when learned
   • Why it matters
   • How to use it next time
5. When nearing limits, rely on context editing to remove stale tool results, but make sure any long-lived knowledge has already been written to memory.

See references/store-vs-keep-checklist.md for common decisions.

Examples

Example: coding agent over a large codebase

• Keep in context: current file, failing test, immediate hypotheses.
• Persist to memory: root cause analysis, architectural decisions, conventions, and next steps.

Example: research agent

• Keep in context: current question, the 1–2 most relevant findings.
• Persist to memory: validated conclusions, citations, and a bibliography pointer.