Agent memory path uses dash instead of dot when agent name contains dots (e.g. hammer.mei → hammer-mei)

Summary

When a subagent's name contains dots (e.g. hammer.mei), Claude Code's injected Persistent Agent Memory system prompt section uses a dash instead of a dot in the memory directory path, creating a mismatch with the actual agent name.

Environment

• Claude Code version: 2.1.121
• OS: macOS
• Agent name: hammer.mei (defined in .claude/agents/hammer.mei.md)
• Agent memory scope: project

What Claude Code injects (actual)

Claude Code automatically appends a "Persistent Agent Memory" section to the system prompt with this path:

/Users/<username>/workspace/myproject/.claude/agent-memory/hammer-mei/

Note: hammer-mei — dot replaced with dash.

What is expected

Per the official docs, the path should be:

.claude/agent-memory/<name-of-agent>/

Which for agent hammer.mei should be:

/Users/<username>/workspace/myproject/.claude/agent-memory/hammer.mei/

Impact

This creates two problems:

1. Empty orphan directory: Claude Code creates .claude/agent-memory/hammer-mei/ (empty), while the real memory lives at .claude/agent-memory/hammer.mei/ (a symlink with actual content).
2. Claude reads the wrong location: The injected system prompt points Claude to the empty hammer-mei/ directory, so it cannot access its existing memories.

Directory state

.claude/agent-memory/
├── hammer-mei/          ← empty directory, created by Claude Code (WRONG)
└── hammer.mei -> /Users/<username>/.tamago/hammer.mei-profile/agents/memory/hammer.mei/
                         ← symlink to actual memory with 30+ memory files (CORRECT)

Observed behavior change

• v2.1.119: No Persistent Agent Memory section was injected into the system prompt (agent memory worked correctly via the agent's own system prompt)
• v2.1.121: The section is now injected with the incorrect hammer-mei path

There is no changelog entry between 2.1.119 and 2.1.121 that documents this change or the dot→dash conversion rule.

Documentation gap

The official docs at https://code.claude.com/docs/en/sub-agents state the path is .claude/agent-memory/<name-of-agent>/ but do not document any name normalization rules (e.g. replacing . with -). If this conversion is intentional, it should be documented.

Possible fixes

1. If intentional: Document the dot→dash normalization rule clearly in the docs
2. If a bug: Use the agent name verbatim (preserving dots) for the memory directory path, so hammer.mei maps to .claude/agent-memory/hammer.mei/