Critical fix
- The OMP hook works with the server runtime (#6367, thanks @flipcyde101, closes #6366 reported by @walynde). When
CLAUDE_MEM_RUNTIMEwasserverorserver-beta, the OMP hook still sent everything to the local worker, so OMP sessions were dropped. The hook now readsCLAUDE_MEM_SERVER_URL,CLAUDE_MEM_SERVER_API_KEYandCLAUDE_MEM_SERVER_PROJECT_ID(or theCLAUDE_MEM_SERVER_BETA_*names; environment variables win oversettings.json) and talks to the server's/v1API directly: session start, tool and reply events, context, and session end. It applies the same rules the worker applies before it saves anything:<private>blocks are stripped, excluded projects andCLAUDE_MEM_SKIP_TOOLSare honored, every prompt is recorded, and the project name comes from git. If the server settings are incomplete, it falls back to the worker as before. - OMP users: run
npx claude-mem install --ide omponce to get this fix. The OMP hook is a copied file, so a plugin update alone does not replace it.
Fixes
- Knowledge queries and corpus requests can no longer hang forever (#6368, thanks @flipcyde101, closes #6364 reported by @soarkit). A stalled Agent SDK stream used to keep the request and its Claude CLI child alive with no end. A knowledge query or prime now stops after 10 minutes and a corpus request after 15, with a 504 error. The corpus stays locked until the timed-out work has really stopped, so it cannot write after a delete or rebuild.
- The "paused saving memories" notice goes away once Codex is saving (#6374, thanks @ryrenz, closes #6372 reported by @raw34). After you switched
CLAUDE_MEM_PROVIDERtocodexfollowing a Claude quota pause, every new session still said capture was paused until the old Claude cooldown ran out. The notice is now recomputed when Codex stores a memory. server api-key create --scopewith no value is refused (#6375, thanks @ryrenz, closes #6373 reported by @soarkit). It used to create a key scoped to the literal stringtrue, which every request then rejected. A bare--scope,--name,--teamor--projectnow prints--<flag> requires a valueand exits 1 without writing anything, on bothcreateandmigrate-scopes.- The Agent Cost Report no longer crashes on observations with no model (#6376, thanks @ZedingZhang, closes #6371 reported by @Realauster).
acr.py rollupraised aTypeErrorwhen one window had both a missing model and a named, unpriced one. - The Docker server image builds again (#6386, closes #6379 reported by @InfoALR). It now uses
node:20-trixie, because tree-sitter 0.26 needs a newer glibc thannode:20ships. - My Work refresh runs on OpenCode 2.x (#6361, refs #6339 reported by @Realauster). OpenCode 2.x rejects the
--pureand--dirflags, so every refresh step exited 1. The step now checksopencode --versionand uses each version's own isolation. 2.0.0 to 2.2.2 and 3.x are refused with a reason, because they cannot keep plugins out. On every version the refresh agent now has all tools denied. - OMP and Hermes read a worker port saved as a number (#6363, thanks @chuenchen309).
{"CLAUDE_MEM_WORKER_PORT": 45678}insettings.jsonused to be ignored, so their requests went to the default port. - The OpenClaw installer no longer deletes a folder it did not install (#6381, thanks @DeryFerd). It ran
rm -rfon whatever install path~/.openclaw/openclaw.jsonnamed. It now replaces a folder only when it is missing, empty, or holds claude-mem's own OpenClaw manifest. Anything else is left alone with a warning and the default extensions folder is used; if that is also taken by something else, the installer stops. - The worker can stop every Claude child it starts (#6385, thanks @DeryFerd). Field compression and the Telegram wrap-up started their Claude process outside the worker's process list, so a worker that exited hard left them running. They are now on that list, and they do not take a concurrency slot.
- A CLI stuck printing can't grow the worker's memory without limit (#6384, thanks @DeryFerd). Output collected from
claude,codex,geminiandopencodeis capped at 128 Mi characters, about 2.5 times the largest real turn measured. Past the cap, the start and end are kept, a warning is logged, and the reply is refused instead of read as a partial answer. Smaller outputs are unchanged. - A transcript record with no end can't grow the worker's memory without limit (#6383, thanks @DeryFerd). A JSONL record over 64 MiB (the largest real record measured was 44 MiB) is skipped, with a warning that names the file, the record's offset and its size. The lines after it are read as normal.
- The unauthenticated
/v1/infono longer shows Redis error text (#6380, thanks @DeryFerd). A queue lane that could not be read used to publish the driver's error, such as a host and port. The reason now stays in the server log.
New
- Opt-in CPU cap for chroma-mcp (#6369, thanks @flipcyde101, partly addresses #6365 reported by @FengJie2004). Set
CLAUDE_MEM_CHROMA_MAX_THREADSto2if chroma-mcp keeps several cores busy. It caps the embedder's threads and, on Linux withtaskset, the CPUs chroma-mcp can use. It is off by default (0), because an always-on cap slowed the first backfill. #6365 stays open. - AIGROW preset for the OpenAI-compatible provider (#6370, thanks @guglielmo-io).
Hosted sync service
- The sync services compare their internal projector credential in constant time (#6382, thanks @DeryFerd). A plain string compare could leak the shared secret through response timing. This changes
services/sync-apiandservices/sync-hub, which are not part of the npm package: it reaches users only once those services are deployed, not through this release.
Tests
- A flaky transcript test now counts the append poll's pokes instead of racing its file check (#6360).
Full changelog: v16.0.2...v16.0.3