agent-commit

Agent Memory Commit

Context

You are an agent that has completed meaningful work. This workflow guides you through capturing high-signal learnings in your agent-scoped memory namespace.

Why agent-scoped memory?

• Isolated knowledge base specific to your expertise
• Prevents polluting project/global scopes with agent-specific patterns
• Enables cross-scope linking for shared insights
• Persistent across sessions

CRITICAL: Dual-Save Pattern

You MUST save memories using BOTH methods:

A. Memory Plugin (Primary - Structured Storage)

echo '{"type":"learning","title":"Brief title","content":"Description","tags":"tag1,tag2"}' | memory write --agent $AGENT_NAME --auto-link

B. Agent Memory Files (Backup - Session Persistence)

Two sub-approaches — use both for maximum redundancy:

B1. Native memory tags (Claude's built-in memory, brief learnings):

<save_memory>
Agent: {agent-name}
Topic: {brief-topic}
Learning: {summary}
</save_memory>

B2. Agent memory directory (file-based, for detailed or structured notes):

Create or append to files under .claude/agent-memory/{agent-name}/:

# Ensure directory exists
mkdir -p .claude/agent-memory/$AGENT_NAME

• MEMORY.md — primary memory file, brief session summaries:

  # Example path: .claude/agent-memory/speckit-explorer/MEMORY.md
  

• Topical files — detailed notes by subject area:

  # Examples:
  # .claude/agent-memory/speckit-explorer/researching-ideas.md
  # .claude/agent-memory/speckit-explorer/patterns.md
  # .claude/agent-memory/typescript-expert/type-system-notes.md
  

Use the Write tool to create/update these files. Keep MEMORY.md under 200 lines — move detail into topical files.

Why both methods?

• Memory plugin (A): Semantic search, graph relationships, structured queries
• Native tags (B1): Survives plugin changes, loads into system prompt automatically
• Directory files (B2): Human-readable, git-tracked, organise by topic, persist across plugin updates
• Redundancy ensures critical knowledge isn't lost