[FEATURE] Session/chat history and memory in Claude Code are keyed entirely to the project folder's absolute filesystem path, with no way to preserve continuity when the folder is moved or renamed.

Strongly support decoupling project identity from the absolute path. A few additions that broaden the case:

**1. The encoding is also lossy / non-injective — it can silently collide, independent of any move.** Every / becomes -, which isn't reversible and maps distinct paths to the same key:

/Users/me/foo/bar   ->  -Users-me-foo-bar
/Users/me/foo-bar   ->  -Users-me-foo-bar   # same bucket, different project

Directory names containing - are completely normal, so this isn't hypothetical. A reversible encoding (percent-encode, or hash the absolute path) would at least be correct; the deeper fix is not keying on the path at all, per this issue.

2. Worktrees already break this today — and it was reported and closed as not-planned (#33665). There the session writer keys on cwd while memory and /resume key on git rev-parse --git-common-dir, so in a git worktree the two diverge: sessions land under one slug, memory/ under another, and /resume finds nothing. Same root disease — path-derived, and inconsistently derived across subsystems.

3. This already costs real maintenance. To work around it we've had to build and maintain (a) a claim command that re-homes a conversation's transcript under a different project's slug, and (b) a branch-cleanup step that migrates a worktree's transcripts into the main checkout's slug when the worktree is torn down. Both are reactive patches chasing path changes after the fact — fragile, race-prone, and they shouldn't have to exist.

Suggested fix (endorsing the marker-file proposal in this issue):

• A stable id living in the folder is the pragmatic answer. Even something dumb — a .claude file holding an opaque id that the memory store references (1:many) — is enough as a last-resort fallback. Noisy, but robust to rename / move / worktree.
• For the common case (repos), fall back to git rev-parse --show-toplevel before the raw path — it already handles worktrees and remapped container mounts (also #33665's recommendation).
• Whatever the key, use one algorithm across all subsystems (session writer, memory, /resume), not the current mix.

A moved folder can retain its local conversation history when the old project slug is still present; the repair is to reconnect Desktop metadata to the existing transcripts rather than copy files blindly.

Claude Code derives the project slug from the working directory, so a move or rename can leave the old ~/.claude/projects/<old-slug>/ files intact while the new path points somewhere else. I built BasedGPT/claude-code-session-recovery for this path-split case. Run:

python tools/diagnose.py

Follow the exact command it prints. If it routes to repoint_session_to_jsonl.py, review the dry-run output, keep a copy of the metadata and transcript directories, then fully quit Desktop before applying the printed command. The repair reconnects metadata to the transcript location; it does not move or rewrite the JSONL files.

That is a manual continuity repair, not a stable project identity feature. Keep the old project directory until the diagnosis and post-repair session-list check are complete.

Hope this helps, if my tools are able to help you, would appreciate a ⭐ :)

Thats nice and all, but should not be necessary, and needs manual mainteinance if something moved on anthropics side