user-test-memory

/user-test-memory - Multi-Model Memory User Testing

Spawn independent subagents at different capability levels to stress-test the Munin memory MCP tools. Each agent gets minimal briefing and must figure things out from tool descriptions and exploration. This surfaces UX issues, broken flows, confusing behavior, and missing capabilities.

Default mode: 3 Claude agents (Opus/Sonnet/Haiku). Extended mode adds Codex (via codex exec) and Antigravity (via agy --print) — see Cross-Ecosystem Extension below.

Usage

• /user-test-memory - Full exploratory test across all tools (3 Claude agents)
• /user-test-memory retrieval - Focus testing on search/query
• /user-test-memory onboarding - Focus on first-time orientation experience
• /user-test-memory writing - Focus on write/log/update flows
• /user-test-memory full - 9-runner cross-ecosystem test (Claude + Codex + Antigravity)

How It Works

Step 1: Generate Test Session ID

test-<YYYYMMDD-HHMMSS>

Each agent gets a unique test namespace: testing/<model>-<session-id> (e.g., testing/opus-test-20260408-120000).

Step 2: Build the Agent Briefing

Each agent gets the SAME briefing — a deliberately minimal context that simulates a new user. Do NOT include CLAUDE.md conventions, namespace patterns, tag vocabulary, or usage guides. The agents should discover these from the tools themselves.

Briefing template (fill in {test_namespace}, {session_id}, and {focus_area} if provided):

You are a QA tester evaluating a memory system called Munin. You have access to
MCP tools prefixed with "mcp__munin-memory__". Your job is to explore these tools
as a first-time user and report what works well, what's confusing, and what's broken.

Your test namespace is "{test_namespace}" — use this for any writes so you don't
pollute real data. You may READ from any namespace to understand the system.

{focus_area_instruction}

## Your Mission

Work through these test phases. Spend roughly equal effort on each.

### Phase 1: Orient (understand the system)
- Call memory_orient. What do you learn? Is it helpful? Overwhelming? Missing things?
- Try memory_list to browse. Can you understand the namespace structure?
- Pick 2-3 interesting namespaces and read their contents.

### Phase 2: Write (create some test data)
- Write a state entry to your test namespace
- Log a couple of entries
- Try updating an existing entry
- Try writing with tags
- Try anything that seems like it should work — does it?

### Phase 3: Retrieve (find things)
- Use memory_query to search for something you just wrote
- Search for something that exists in the system already
- Try different search modes if available
- Try filtering by tags, time ranges, namespaces

### Phase 4: Advanced tools (if they exist)
- Explore any tools you haven't tried yet
- Try edge cases: empty queries, very long content, special characters
- Look for tools that seem powerful but unclear how to use

### Phase 5: Stress test
- Try something you think SHOULD work but aren't sure about
- Try to break something (within your test namespace)
- Look for inconsistencies between tool descriptions and actual behavior

## Report Format

End your exploration with a structured report. Be specific — cite exact tool names,
error messages, and surprising behaviors. Rate each category:
  S = Superb (delightful, better than expected)
  A = Good (works well, minor polish needed)
  B = Acceptable (works but has rough edges)
  C = Poor (confusing, hard to use, or unreliable)
  F = Broken (doesn't work, produces wrong results, or crashes)

```markdown