thedotmack/claude-mem v15.18.5

on GitHub

Critical fixes

• The tree-sitter CLI installs on networks where its download script keeps failing (#6110, thanks @rigel905, closes #6109). On some networks tree-sitter-cli's own install.js had its request reset every time, or exited 0 on a truncated file. The pinned SHA-256 check correctly refused the file, but nothing recovered after that, so smart_search, smart_outline, smart_unfold and the File Read Gate stayed off on every worker start, and npx claude-mem repair aborted. Now, when the script fails or leaves anything but the pinned executable, claude-mem fetches the same release asset directly, once, inside the same install time budget, and runs it only after it unpacks to the same pinned SHA-256. The script still runs first, so HTTPS_PROXY keeps working. A script that used up the whole budget gets no second attempt.
• A corrupt Chroma collection is dropped once, not once per parallel backfill (#6115, thanks @mohitdixit02 and @0xABCD01, closes #6113). When the HNSW segment of a large collection was corrupt, every parallel per-project backfill that hit it passed the "drop only once" check before the first delete finished, and queued its own delete against chroma-mcp. In #6113 that ran 19 drops at once and pushed chroma-mcp to a 312 GB memory footprint, stalling a 32 GB Mac. Callers now join the one drop already running for that collection and share its outcome. A failed drop costs one attempt from the retry budget, not one per caller, and different collections still drop independently. The remaining asks from that report (a memory bound on chroma-mcp, stopping writes when a drop can't be confirmed, measuring footprint on macOS) are tracked in #3610.

Fixes

• --help prints help instead of running the command (#6142, thanks @liuyinchu). npx claude-mem mcp --help used to start the stdio server, and status, doctor and hook entered their runtime paths; t3code and its aliases rejected the flag. -h and --help are now handled before dispatch for these commands. Plain mcp still starts the server.

Docs

• OmO Native recall guide (#6087, thanks @mschoi0411). How to connect OmO Native to the local worker through the existing stdio MCP server: tool discovery, recall by project and observation ID, troubleshooting and disconnect.