memory

Memory System

You have access to project-specific decision memory that evolves over time. Memory changes how you act, not what you know.

Memory is NOT context. Facts about the project (stack, structure, files) belong in context/project.md. Memory stores decision rules — what to do, what to avoid, what works, what doesn't.

On Startup — READ

1. Check if .claude/transformers/memory/long-term/index.md exists

   • If YES → read it. These are decision rules from past work. Apply relevant ones.
   • If NO → no memory yet. Proceed normally.

2. Check if .claude/transformers/memory/temp.md exists

   • If YES → read it. These are decision rules from earlier in this conversation.
   • If NO → proceed normally.

3. If a rule in the index is relevant to your current task, read the full category file for the reasoning behind it.

During Work — LEARN

When you had to figure something out the hard way — that's a memory candidate.

The test: Would knowing this change how an agent acts next time? If yes, save it. If no, skip it.

Tag it in your response to the orchestrator:

[MEMORY] brief decision rule — why

If you're a standalone command (no orchestrator), spawn Scribe directly:

"Append to .claude/transformers/memory/temp.md: ## [timestamp] [your-name] — [title]\n[decision rule and why]"

For rules that are clearly permanent, tag as:

[LONG-TERM] category: decision rule

What IS Memory

• Decision rules: "always do X before Y in this project"
• Strategy insights: "skip agent X for Y-type work — wastes tokens"
• Failure patterns: "when you see X symptom, root cause is usually Y"
• Workflow rules: "pull before push — CI modifies remote"

What is NOT Memory

• Project facts (stack, structure, files) → project-context.md
• Task progress → status.md
• Activity history → activity.log
• Test/build results → reports
• Anything discoverable from code or docs

Full memory system rules: @${CLAUDE_PLUGIN_ROOT}/knowledge/memory-system.md