Symptom

Windows users running mm init don't get Claude Code / Codex memory directories auto-detected by the wizard's "Provider memory folders" step. Any such path they add manually is classified under the generic user category instead of claude-memory / claude-plans / codex, so the preset namespace rules added in #312 (claude:{ancestor:1}, claude-plans, codex) don't get applied either.

Silent failure: no error, no warning — the wizard just doesn't find the directories it would find on macOS/Linux.

Root cause

categorize_memory_dir at packages/memtomem/src/memtomem/config.py:1040-1064 uses regex patterns that match forward slashes only:

_PROVIDER_CATEGORY_PATTERNS = (
    ("claude-memory", re.compile(r"/\.claude/projects/[^/]+/memory/?$")),
    ("claude-plans",  re.compile(r"/\.claude/plans/?$")),
    ("codex",         re.compile(r"/\.codex/memories/?$")),
)

On Windows, str(Path(...)) returns backslash-separated paths (C:\Users\foo\.claude\projects\bar\memory), so none of the patterns match and the function falls through to return "user" (line 1064). The docstring at lines 1056-1058 already acknowledges this gap:

Uses forward-slash regex, so on Windows a backslash-normalised path will fall through to "user" until a future path-sep-agnostic pass lands.

This issue is that "future pass".

Scope

• Normalise the path-separator before matching (e.g., Path(path).as_posix() or str(path).replace("\\", "/")).
• Preserve existing POSIX-path behaviour — no regression for macOS/Linux users.
• Independent of RFC RFC: nested provider/product hierarchy for memory_dir categorization #304 vocabulary decisions. The category names and the axis stay exactly as they are today; only the matching mechanism changes.
• The vocabulary lock from fix(config): lock provider category vocabulary (RFC #304 prep) #313 (_VALID_PROVIDER_CATEGORIES) stays untouched — this fix does not modify the locked set.

Testing notes

CI (.github/workflows/ci.yml) runs on ubuntu-latest only — no Windows runner. A fix therefore needs platform-agnostic tests that exercise the backslash code path without a Windows host.

Preferred fixture pattern: hand-constructed backslash strings passed directly to categorize_memory_dir, e.g.

assert categorize_memory_dir(r"C:\Users\foo\.claude\projects\abc\memory") == "claude-memory"

Raw strings (r"...") avoid backslash-escape confusion. This exercises the matching logic directly without touching platform-specific path objects. pathlib.PureWindowsPath also works if a fixture prefers path objects, but the raw-string form is the simpler default.

Acceptance criteria

• Windows-style paths (C:\Users\…\.claude\projects\<id>\memory, …\.claude\plans, …\.codex\memories) classify into the correct provider category.
• Existing POSIX-path tests continue to pass — no regression in TestCategorizeMemoryDir / TestDetectProviderDirsRoundtrip.
• New tests cover at least one Windows-style path per provider category plus the user fallback, using raw-string backslash fixtures (or PureWindowsPath) so they run on Linux CI.
• Docstring at config.py:1056-1058 updated to reflect that the gap is closed.
• Edge case consideration: document behaviour for mixed separators (e.g., C:\Users\foo/.claude/plans) and UNC paths (\\server\share\...) — may be out of scope; contributor's call with rationale in the PR body.

Related

• fix(config): lock provider category vocabulary (RFC #304 prep) #313 — vocabulary lock (unchanged by this fix; only the match mechanism changes).
• feat(cli): wizard appends matching namespace rules for accepted providers (closes #296) #312 — wizard preset namespace rules (directly benefits — Windows users then get provider rules applied).
• Windows: test components fixture leaves SQLite/embedder handles open, blocking pytest cleanup #206 — separate Windows test-fixture cleanup issue; different gap, no overlap.
• RFC RFC: nested provider/product hierarchy for memory_dir categorization #304 — axis/wire format decisions are independent; this fix is not blocked on them.