thedotmack/claude-mem v13.29.0

on GitHub

The agent keeps its to-do list in claude-mem

Every session now opens with a work-state section ahead of memory: a rule telling the agent to use claude-mem's work_state_write and work_state_read tools as its canonical to-do list, then whatever is still open in the project. Claude Code gives Claude 5 models no native to-do tool, so until now nothing recorded what was in progress or what was left when a session ended. The rest of the release adds an openai-compatible provider with presets, a Codex subscription provider, Kimi Code and Oh My Pi support, an opt-in quota fallback, and fixes across capture, search, the worker and Windows. Several defaults changed; see Upgrade notes.

Work state

• Two MCP tools. work_state_write appends one entry to a named list (fields capped at 2,000 characters as JSON) and replies with what is still open in that list. work_state_read returns every open item, or one list, and adds done and dropped items when includeClosed is true. (#4340)
• Lists, items and state. An entry with a task field is a to-do item with a status of todo, doing, done or dropped; any other field is state on the list. The latest value of each key wins, null clears a key, and "status": "done" closes the list. (#4340)
• SessionStart leads with it. Context opens with the rule and every open item, each with how long ago it last changed, capped at 3,000 characters (806 when nothing is open); memory is fitted into what remains of the 10,000-character limit. The welcome hint for a project with no observations gets the same lead; the colored terminal preview does not. (#4340)
• Scoped like memory. Writes land under the checkout's primary project key and reads cover every key the checkout reads, so a list started in the main checkout also shows in its worktrees. Writes from projects matched by CLAUDE_MEM_EXCLUDED_PROJECTS are not saved. (#4340)
• Kept in the database, not the repo. Entries go to a new append-only work_state_entries table (schema v61), so lists never conflict on merge but are not in git, and reads and writes need the worker. The server runtime's SessionStart does not read this table yet. (#4340)

Upgrade notes

• Observer deadline raised to 3 minutes. CLAUDE_MEM_LLM_TIMEOUT_MS now defaults to 180000 (was 30000), and the new CLAUDE_MEM_FIELD_OPTIMIZE_TIMEOUT_MS for oversized-field condensation (a fixed 30 seconds before) defaults to the same. A settings.json holding exactly the old 30000 is moved once; any other value is kept. (#4278, #4287, #4136)
• New OpenRouter default model. OpenRouter retired xiaomi/mimo-v2-flash:free, so the default is now cohere/north-mini-code:free, and a stored CLAUDE_MEM_OPENROUTER_MODEL equal to the retired id is rewritten once when the base URL is blank or openrouter.ai. A retired or unknown model now fails with a model_unavailable error that names the setting. (#3662)
• Subagent observations stay out of SessionStart. CLAUDE_MEM_CONTEXT_MAIN_AGENT_ONLY (default true) leaves rows that carry both an agent id and an agent type out of the SessionStart window, in the worker and the server runtime. Search, timeline and get_observations still return them; set it to false for the old behavior. (#3310, #4290)
• Hooks never block a session. No claude-mem hook exits 2 any more, so an unexpected error or a worker outage no longer drops your prompt, denies a tool call or re-wakes Stop; the session continues without memory. After CLAUDE_MEM_HOOK_FAIL_LOUD_THRESHOLD (default 3) unreachable-worker hooks in a row you get one notice per session to run npx claude-mem restart, and Claude Code hooks other than Setup exit 0 even when the plugin scripts are missing. (#2892, #3168, #3431)
• UserPromptSubmit gets 15 seconds in Claude Code. The hook timeout drops from 60 to 15 seconds, and session init spends one budget, CLAUDE_MEM_SESSION_INIT_TIMEOUT_MS (10 seconds; 7 on Windows, where longer saved values are cut to 7), across the whole worker round-trip. (#3440, #4299)
• The worker refuses unknown host names. As a DNS-rebinding guard, a request whose Host is a DNS name other than localhost, *.localhost, host.docker.internal, CLAUDE_MEM_WORKER_HOST, the CLAUDE_MEM_PUBLIC_URL host or an allowed origin gets a 403; IP addresses still work. If you open the viewer through another host name (a .local name, a reverse proxy), add its origin to the new CLAUDE_MEM_ALLOWED_ORIGINS (comma-separated) and restart the worker. (#3514)
• Antigravity users: re-run install. Earlier installs wrote ~/.gemini/config/hooks.json in a shape agy ignores, so nothing was captured. Run npx claude-mem install --ide antigravity to rewrite it in agy's named-hook schema; npx claude-mem antigravity-cli status reports entries left to migrate. (#4274)

Providers and gateways

• New openai-compatible provider. CLAUDE_MEM_PROVIDER=openai-compatible runs the observer against any OpenAI /chat/completions endpoint, configured with CLAUDE_MEM_OPENAI_COMPAT_PRESET, CLAUDE_MEM_OPENAI_COMPAT_BASE_URL, CLAUDE_MEM_OPENAI_COMPAT_MODEL and CLAUDE_MEM_OPENAI_COMPAT_API_KEY. Endpoints on localhost or a private network need no key, and a selected but half-configured provider falls back to Claude. (#3942)
• Presets. nvidia-nim, deepseek, groq, together, ollama, lmstudio and vllm (#3942), orcarouter (#3581), opencode-go and opencode-zen (#3623), minimax and minimax-cn (#3376), api-route (#4325) and opper (#4333). Model ids are passed verbatim, and a leading <think>…</think> block is stripped from replies so inline reasoning is not stored (#3376).
• Endpoint settings in the viewer. A new "OpenAI-compatible endpoint (BYOK)" provider option edits the preset, base URL and model (the key stays in settings.json or .env), and the OpenRouter block shows CLAUDE_MEM_OPENROUTER_BASE_URL, read-only while it points at the claude-mem observer. The settings API rejects base URLs that are not http(s). (#3433)
• Codex subscription provider. CLAUDE_MEM_PROVIDER=codex, or npx claude-mem install --provider codex [--model <id>], runs the observer through the Codex CLI app-server on your ChatGPT subscription; run codex login first. CLAUDE_MEM_CODEX_MODEL (empty uses Codex's default) and CLAUDE_MEM_CODEX_REASONING_EFFORT (default low) are optional, and CLAUDE_MEM_CODEX_PATH (default codex) can be set only in settings.json or the environment. (#3882, #4216, #4318)
• Codex throughput. Codex runs up to CLAUDE_MEM_CODEX_MAX_CONCURRENT_AGENTS requests at once (default 2) and sends consecutive queued observations in one request, up to CLAUDE_MEM_CODEX_OBSERVATION_BATCH_SIZE items (default 8) and CLAUDE_MEM_CODEX_OBSERVATION_BATCH_MAX_CHARS characters (default 32000). The viewer does not expose these three; set them in settings.json. (#4260, #4261)
• Codex failures are classified. A request Codex refuses the same way every time becomes a setup error with a remedy instead of an endless retry, app-server internal errors (-32603, -32700) are transient, and an empty reply goes through the normal skip handling. On Windows the default codex command launches the codex.cmd shim. (#4222, #4316, #4323)
• Several keys per provider. CLAUDE_MEM_GEMINI_API_KEYS, CLAUDE_MEM_OPENROUTER_API_KEYS and CLAUDE_MEM_OPENAI_COMPAT_API_KEYS take extra keys, separated by commas or newlines. A key that hits a rate limit, a spent quota or an auth failure is set aside for 1 minute, 30 minutes or 6 hours and the next key is used; the cmem.ai gateway never uses a pool. (#3941)
• Quota fallback provider (opt-in). CLAUDE_MEM_QUOTA_FALLBACK_PROVIDER (claude, gemini, openrouter or openai-compatible; empty, meaning off, by default) takes observer work while the selected provider is in a quota cooldown and hands it back once the primary answers its probe. CLAUDE_MEM_QUOTA_FALLBACK_MODEL picks the model for a Claude fallback, both are in the viewer, and authentication failures are never routed around. (#4268, #4306)
• cmem.ai trial-expiry fallback works again. Since v13.24.11 a terminal gateway rejection never moved memory to your Anthropic plan. Now allowance_exhausted, key_invalid, subscription_inactive or any gateway 401/403 switches to Claude at once and resumes the buffered work there, the SessionStart notice relays the gateway's own message and link, and unattended resumes on the gateway are capped at 3 in a row. (#4276, #4273, #4306)
• OpenRouter request options. CLAUDE_MEM_OPENROUTER_REASONING_EFFORT (none, minimal, low, medium or high; empty sends nothing) sets reasoning for requests to openrouter.ai only (#3001). CLAUDE_MEM_OPENROUTER_EXTRA_BODY (empty by default) merges a JSON object into OpenRouter-provider requests, never to the cmem.ai gateway, and cannot override model, messages, stream or the token caps (#3040).
• Gateways that stream by default. Requests to openrouter.ai and custom base URLs send stream: false in the worker and the server runtime, so a gateway that streams unless told otherwise no longer fails every observation (#3668, #4317). A litellm "Unable to get json response" reply is retried (#3263), and an OpenRouter reply cut at the output cap is logged with the cap and request id (#3620).
• Observer output cap and system prompt. The new CLAUDE_MEM_OBSERVER_MAX_OUTPUT_TOKENS (default 4096) sets the output cap on OpenRouter, openai-compatible, cmem.ai gateway and Gemini observer requests, which now carry the observation schema in the system message (systemInstruction on Gemini). (#3868)
• Gemini replies and throttles. The Gemini provider reads the answer parts of a reply instead of a leading reasoning part, which had stored nothing (#4090). A 429 that names only a per-minute limit is retried after Google's delay instead of pausing capture for 30 minutes, read from the structured error details (#4089, #4243).
• Daily caps and region refusals. A per-day 429, such as OpenRouter's free-models-per-day, counts as a spent daily allowance in the worker and the server runtime (#4089, #4308). Gemini's "User location is not supported" pauses capture as location_unsupported with advice to choose another CLAUDE_MEM_PROVIDER, without rotating through the key pool (#4147, #4308).
• HTTP provider errors. On openai-compatible endpoints a context-length refusal recycles the observer generation instead of dropping the batch, a per-minute 429 that links to a billing page stays a rate limit, and a key pool waits for a briefly throttled key instead of arming the 30-minute breaker (#4317). An OpenRouter or openai-compatible request that gets no response names the error code and host, with a Local Network permission hint for hosts on your network (#4115, #4320).

New hosts and harnesses

• Kimi Code CLI. npx claude-mem install --ide kimi, or npx claude-mem kimi install|status|uninstall, writes hooks to $KIMI_CODE_HOME/config.toml and the MCP server to $KIMI_CODE_HOME/mcp.json (default ~/.kimi-code). Memory is injected on a session's first prompt and after compaction, tool output and failed calls are captured, and Kimi's bookkeeping tools (to-do, task and cron) are skipped. (#3676, #3547)
• OMP (Oh My Pi). npx claude-mem install --ide omp installs a hook at ~/.omp/agent/hooks/pre/claude-mem.ts (under $PI_CODING_AGENT_DIR when set) that records sessions, every prompt, tool results and summaries and injects memory before model calls. Its worker requests time out after 5 seconds. (#3556, #4315)
• DeepSeek Harness transcripts (opt-in). The transcript watcher can tail zstd-compressed *.jsonl.zstd session logs, and the repository's transcript-watch.example.json has a dsh schema and watch entry for ~/.dsh/sessions to copy into ~/.claude-mem/transcript-watch.json. (#3691)
• OpenCode plugin loads and files sessions correctly. On OpenCode 1.18 the plugin failed with "Plugin export is not a function"; it now loads, tags sessions opencode, keys them by checkout like other hosts, and records the real prompt (#3803). Memory context is added to each request's system prompt, and captures refused while the worker boots are retried for about 10 seconds (#3208, #4091).
• OpenCode configuration. Install registers the claude-mem MCP server in opencode.json (#3621), stops writing memory into the global ~/.config/opencode/AGENTS.md and removes a block an older install left there, and a failed context fetch is retried after a minute instead of on every system prompt (#4314).
• Codex hooks. Codex hook events keep tool_use_id, agent_id and agent_type, so dedup and the subagent settings apply (#4329). Codex's own helper prompts (task titles, memory consolidation, onboarding) no longer become sessions (#2810), and worktrees at ~/.codex/worktrees/<id>/<repo> file memory under <repo> instead of <repo>/<repo> (#3643).
• Codex subagent capture (opt-in). Codex hooks do not fire for subagent threads, so set CLAUDE_MEM_CODEX_SUBAGENT_INGESTION=true (default false) for a transcript watch scoped to subagent rollouts. CLAUDE_MEM_SKIP_SUBAGENT_OBSERVATIONS=true still wins. (#3655, #4321)
• Antigravity, Qwen Code and the claude alias. Sources named agy, antigravity or antigravity-* are recorded as antigravity-cli with their own viewer badge (#4147), and --ide claude and hook claude <event> behave like claude-code (#2835). A host-sent submitted_prompt is stored as the prompt, and an empty one no longer stores a [media prompt] row for every tool round in Qwen Code (#4224).
• Grok Bot brainbeat webhook (opt-in). Set CLAUDE_MEM_GROK_BOT_WEBHOOK_URL and claude-mem POSTs each observation that matches the Grok Bot awareness triggers to it, sending CLAUDE_MEM_GROK_BOT_WEBHOOK_SECRET as a header when set. It is off while the URL is empty. (#3846)

Memory, context and search

• Named environments. CLAUDE_MEM_PROJECT_ENVIRONMENTS (default [], off) takes a JSON array of {"name", "patterns"}, and folders matching an environment's globs share one project named after it. npx claude-mem project merge <from> <into> [--dry-run] folds one project's memory into another without rewriting rows, runs inside the worker when one is up, and reaches other devices through cloud sync. (#2737, #4297)
• Stable project names. CLAUDE_MEM_PROJECT_NAME_SOURCE=git-remote (default path) names projects by the org/repo slug of the git origin remote (#2827), and outside git an empty .claude-mem-project file (or .claude-mem.json) at the project root gives every subfolder the root's name (#3311). Search covers every key a checkout has used, so memory filed under an older name is still found (#4303).
• Case-insensitive project keys. Checkouts whose folder names differ only in case share one memory bucket in context, search and folder CLAUDE.md, including every Chroma path. (#3536, #4301)
• Worktree and submodule memory. The startup sweep that folds merged worktrees into their parent project works again and adopts observations from deleted worktree checkouts. A session inside a git submodule uses a <superproject>/<path> key and reads the superproject's memory. (#3525, #4291)
• Opt-in ACT-R ranking. Setting CLAUDE_MEM_REINFORCE_ALPHA (default 0, off) above 0 lets older observations re-confirmed on later days climb into the SessionStart window; the newest quarter of the window always stays by recency. (#3414)
• Opt-in near-duplicate dedup. With CLAUDE_MEM_DEDUP_ENABLED=true (default false), an observation whose normalized title matches one in the same project, platform and agent scope is merged into it instead of stored again. Near matches are only listed at GET /api/dedup/candidates, and POST /api/dedup/scan backfills. (#3063, #4294)
• Import Claude Code auto-memory. npx claude-mem memory ingest [--source <dir> | --all] [--dry-run] stores Claude Code's auto-memory files (~/.claude/projects/<encoded-cwd>/memory/*.md) as observations with no model call, from the checkout you run it in. It skips MEMORY.md, symlinked notes and notes over 64 KiB, and dedupes re-runs. (#2829, #4322, #4307)
• Keyword matching. Multi-word queries now require every word anywhere instead of one exact phrase, punctuation-only tokens are dropped (#4180), and mixed Latin and CJK queries match term by term (#4148). Searches with only obs_type or concepts return rows instead of a 400 (#3259).
• SQLite fills gaps in semantic search. A category Chroma returns nothing for (common with CJK queries or a tight date window) is refilled from SQLite FTS5 (#3173, #3285), and a query-filtered build_corpus with CLAUDE_MEM_CHROMA_ENABLED=false falls back to full-text search (#4285).
• Results and ordering. Observation results are grouped under day headers (#3693), date-ordered searches merge keyword matches before cutting to the limit (#4174), timeline counts depth correctly when rows share the anchor's timestamp (#4104), and broad project filters on large Chroma collections are applied client-side (#3675).
• Folder and file lookups. Folder lookups match files stored project-relative and keep exact file matches Chroma did not rank, so opt-in folder CLAUDE.md files (CLAUDE_MEM_FOLDER_CLAUDEMD_ENABLED) are no longer empty (#3116). Search routes take projects as a comma-separated list or a repeated key and answer other shapes with 400 INVALID_PROJECTS (#4304).
• Knowledge corpora. build_corpus keeps the dateStart / dateEnd you pass and is no longer limited to 90 days of Chroma matches, and rebuild_corpus refuses a rebuild that would keep half or fewer of a corpus's observations unless called with force: true. (#4168)
• SessionStart from every harness (opt-in). CLAUDE_MEM_SESSION_START_INCLUDE_ALL_SOURCES=true (default false), or Include all sources at session start in the viewer, loads recent observations from every harness for the project, not only the one starting. (#4251)
• SessionStart fixes. The last session summary shows again when it shares a timestamp with the newest observation (#4103), the terminal preview shows the observations the model received (#4254), and a resumed session such as claude --resume is set back to active (#4084).
• New /handoff skill. It writes a HANDOFF.md with the goal, current state, files in play, failed attempts, next steps and memory pointers, so a fresh session can continue. (#3748)

Capture

• Failed tool calls are observations. Claude Code sends a failed tool call (a non-zero Bash exit, or a call you interrupted) to PostToolUseFailure, which claude-mem now registers; the stored tool use records the error and whether it was interrupted. (#4339)
• Skip shell commands by pattern. CLAUDE_MEM_SKIP_BASH_PATTERNS (empty by default) is a regular expression tested against each Claude Code Bash or Codex exec_command command, for example ^(ls|cat|pwd)\b; a match skips the observation. (#3562)
• Skip subagent observations (opt-in). CLAUDE_MEM_SKIP_SUBAGENT_OBSERVATIONS=true skips every subagent tool call and CLAUDE_MEM_SKIP_AGENT_TYPES skips listed agent types (for example workflow-subagent,Explore), in the hook before any provider request. Both are off by default. (#2741)
• Turn off the tool hooks from the environment. CLAUDE_MEM_DISABLE_TOOL_HOOKS=1 makes the PreToolUse and PostToolUse hooks exit at once, and CLAUDE_MEM_DISABLE_OBSERVATION=1 or CLAUDE_MEM_DISABLE_FILE_CONTEXT=1 turns off one of them. They are read from the hook's environment, not from ~/.claude-mem/settings.json. (#3316)
• Record Claude Code advisor calls (opt-in). With CLAUDE_MEM_CAPTURE_ADVISOR_CALLS=true (default false), the Stop hook stores each advisor call's advice, model and prompt locally at no token cost, readable at GET /api/advisor-calls and never synced. Worker runtime only. (#3165)
• Summaries survive a session that ends mid-tool-call. An empty last assistant message no longer skips the summary: the hook reads the transcript, or names the last tools used. (#3184)
• Less observer noise. The default code mode tells the observer to record a fact once instead of re-confirming it during reviews, retries and polling (#3352), and a prompt that queues no work no longer starts an observer request (#3467).
• Faster hooks. The SessionStart worker start runs async while the context hook stays synchronous (#3401), hooks use a valid CLAUDE_PLUGIN_ROOT instead of scanning the plugin cache (#3470), and hook processes no longer read viewer assets at import (#3667).

Viewer and settings

• Sessions view. The viewer has Timeline and Sessions tabs. Each session opens a page you can filter by observation category, and you can delete a whole session unless it is still running or holds memories synced from another device. (#3458)
• Delete memories from the viewer. Observation and summary cards have a delete button with a confirmation, deletes are cloud-sync safe, and other open tabs drop the row live. Browser DELETEs are accepted only from the viewer's own page or an origin in CLAUDE_MEM_ALLOWED_ORIGINS. (#2925, #4288)
• Settings saves. A save is validated only on what it changes (#4320), never writes a value that only echoes an environment variable into settings.json (#4324), and shows the server's reason when it fails (#3474); a body-less POST /api/settings gets a 400 (#3632).
• settings.json handling. The installer moves a corrupt ~/.claude-mem/settings.json to settings.json.corrupt-<epoch-ms> instead of wiping it, and the worker and viewer refuse to write over one (#3472). When the file's env block holds CLAUDE_MEM_* keys, they win over stale root copies, which the next save removes (#4286).
• Links behind a port-forward. CLAUDE_MEM_PUBLIC_URL (empty by default) sets the browser-reachable base for viewer links in the welcome hint, SessionStart, the per-prompt banner and health notices, for remote sandboxes and dev containers. (#3323)
• Theme colors before the theme loads. The fallback colors match the dark theme, and source badges use theme tokens in both themes. (#4328)

Installer and CLI

• Fresh installs load in Claude Code. The npm package ships .claude-plugin/marketplace.json, so a clean install no longer ends with "Marketplace thedotmack failed to load: cache-miss". (#3441)
• Installs that used to abort. npm 11.16+ no longer aborts with EALLOWSCRIPTS (#3258). Worker liveness comes from the PID file, so slow refusals (WSL2) and a port held by another process no longer stop an install before sign-in (#4119), and a port taken again mid-install warns instead of aborting (#4289).
• tree-sitter CLI is provisioned. The installer runs tree-sitter-cli's install step when the binary is missing, so smart_search and smart_outline work on installs that suppress lifecycle scripts, and doctor gains a "tree-sitter CLI" row. (#3254)
• Old plugin versions are pruned. Install and update remove superseded plugin cache versions, keeping the newest 2, the running worker's and the one Claude Code has registered; npx claude-mem prune [--dry-run] [--keep <n>] runs it on demand. (#4106)
• update keeps your provider setup. A non-interactive update keeps openai-compatible setups, and Gemini or OpenRouter keys that live only in a key pool or ~/.claude-mem/.env. (#4312)
• doctor and status find cache installs. The CLI resolves the plugin root the way the worker does and prefers the newest complete copy, so working cache installs are no longer reported missing. (#3535)
• Externally managed workers. CLAUDE_MEM_WORKER_AUTOSTART=false (default true) makes hooks, the MCP server, start and the installer use a running worker but never launch, stop or recycle one. (#2828, #4289)
• Smaller fixes. Marketplace installs record the install marker instead of printing "runtime not yet set up" at every Setup (#3113), /api/import reports rejected rows instead of aborting the batch (#4051), and CMEM Pro trial copy says "up to 14 days" (#4265).

Server runtime

• SessionStart reads the shared server. With CLAUDE_MEM_RUNTIME=server, SessionStart renders context from the server's rows in one /v1/context read sized to the host's hook limit, with short display refs that point to observation_search, and start no longer probes or spawns a local worker. (#3227, #4112, #4290, #2867)
• Server API. /v1/context no longer needs a query: without one it returns the newest observations (50 by default, at most 200) (#4079). The observation_add MCP tool now sends content, which /v1/memories validates; every call had failed with a 400 (#4143).
• Generation. The CLAUDE_MEM_SERVER_GENERATION_CONCURRENCY environment variable (unset means 1, clamped at 64) sets parallel jobs per lane (#4078), summaries read the session's first and last 500 events (#4137), provider calls time out after 10 minutes (#4123), and empty replies fail at once instead of being retried (#3368).
• Custom generation provider. CLAUDE_MEM_SERVER_PROVIDER=custom with CLAUDE_MEM_CUSTOM_PROVIDER_MODULE=<absolute path> loads your module's createProvider(helpers) factory at server start. (#3228)

Reliability: worker and hooks

• A wedged worker is reclaimed, on evidence only. A claude-mem worker that holds its port and PID file but stops answering health is replaced once it is older than CLAUDE_MEM_WEDGED_WORKER_UPTIME_S (default 300) and its identity checks pass; a foreign process on the port is never touched. (#4129)
• Port handling. Hooks prove the port is free before spawning (#3219) and wait until a killed worker's port can be bound (#3416). An unbindable port is reported as a boot failure (#4299), and a worker that dies before binding exits 78 and logs its last boot step (#3558).
• PID files. A worker.pid whose process is alive but never healthy is removed once the port is free, Bun is also found through BUN_INSTALL (#3306), and PID namespaces such as bwrap --unshare-pid no longer delete a healthy worker's PID file (#4158).
• Self-healing. When Claude Code's auto-updater leaves the claude CLI unspawnable, the worker restarts itself, at most 3 times an hour (#3291). The daemon moves its working directory to the data directory, ending EPERM spawns from a launch folder it cannot re-enter (#3252), and an unwritable process registry degrades to memory with one warning instead of crashing (#4142).
• Hook output is flushed. Hooks wait for stdout to finish before exiting, so a large SessionStart context read through a pipe is no longer cut off. (#4331)
• Full-text index bloat. Index triggers fire only when indexed columns change, the unused user_prompts_fts table is dropped, and existing bloat is merged away once, in small background steps starting a minute after the worker starts. (#3284)
• Buffered work resumes. Work paused by an observer deadline resumes with backoff without waiting for another prompt (#4273), and a summary that arrives during an idle shutdown goes to a new observer (#3421).
• Diagnostics. /api/health fills ai.lastInteraction (#3197), stopped workers are recognized under Bun and Node (#3447), malformed input and circular log data no longer overflow the stack (#3264), and object log context prints as JSON (#4157).

Reliability: observer and quota

• Budgets follow the model's context window. The observer resolves its model's window (or CLAUDE_MEM_OBSERVER_CONTEXT_WINDOW), retires a conversation at half of it using measured tokens, caps each tool field at a tenth, and stubs out old tool payloads on HTTP providers (#3625, #2957, #3151). A context-length refusal recycles the conversation and keeps the batch, and retry loops no longer grow history (#3625, #3479).
• Replies that are not answers keep their batch. Prose or empty replies get one retry in a fresh conversation (#3624), Claude CLI transport-error text is retried (#3460), drifted XML tags are corrected instead of copied (#3475), and capitalized tags from local models parse (#4113).
• Field condensation. Screenshots (MCP image blocks and data:image URLs) are replaced with a marker before condensation, a field over half the context window is truncated without a call (#4327), and a condense reply cut at the token limit is not stored (#4332).
• Claude observer. The empty tool list reaches the Claude CLI as --tools= (#3615), CLAUDE_CODE_TMPDIR is kept (#4162), the default macOS profile can refresh its token (#4153), a signed-out CLI shows in /api/health and the SessionStart banner (#4154), and the working directory is created before spawn (#4117).
• Quota state per Claude account. Rate-limit readings and cooldowns are tagged with the account that recorded them, so switching with CLAUDE_MEM_CLAUDE_CONFIG_DIR no longer inherits a pause (#4272, #4296). Queues held by a cooldown resume when it ends (#4110).
• Quota windows. The seven_day_overage_included window is tracked and pauses work only when a request is refused on it (#4132, #4293), the five-hour and seven-day windows refresh from the CLI's live figures (#4226), and /api/health marks expired windows and stores reset times in one unit (#4072, #4155, #4263).
• Accurate outage banners. A quota outage older than 30 minutes shows as a last-known-state note (#4085), expired cooldowns are not shown as active (#4116), and the banner stays out of the observer's own briefing (#4225).
• No lost or misfiled observations. The stall timer pauses while a reply is stored, so a slow store no longer stores a batch twice (#4077), and stored IDs stay paired with their observations when a batch holds an untitled one (#4294).

Reliability: Chroma, transcripts and cloud sync

• Corrupt collections are rebuilt. Only a confirmed HNSW segment error on 2 batches drops a collection, which is then rebuilt from SQLite; a failed drop is retried and a chroma-mcp hung on a read is restarted (#3203, #4295). doctor reports chroma-mcp exits, the prewarm breaker and dropped collections (#3393).
• Slow writes no longer kill chroma-mcp. A timed-out request fails only that call and leaves the row pending, writes get CLAUDE_MEM_CHROMA_MUTATION_TIMEOUT_MS (default 600000), and NUL bytes are replaced before writes. (#3571)
• Chroma versions and settings. chromadb is pinned to 1.5.9 (#3384), and the local store now records its writer epoch, so a future downgrade reports vector search unavailable instead of crashing on a newer store (#3466). CLAUDE_MEM_CHROMA_EMBEDDING_FUNCTION (default default) picks the embedding function for new collections; the API-backed ones send memory text to their vendor (#3422).
• Backfill. Title-only observations are indexed (#4264), partial failures finish as rows_pending and retry only the failed rows (#4279), and shutdown stops a backfill instead of reporting it complete (#4120).
• Prewarm and locks. A failing uvx prewarm backs off after 5 failures and stops after 20, with UV_LINK_MODE=copy on Windows (#4111), and a writer lock left by a reused PID is reaped (#4239).
• Transcript turns are kept. Transcript-watched sessions such as Codex rollouts record each user turn, so their observations are no longer lost (#3654). A turn the worker did not record is retried from its own line or zstd frame, and records caught mid-write are neither dropped nor duplicated (#4313, #4192, #4193).
• Transcript watcher. A turn with no known working directory waits for one instead of being filed under ~/.claude-mem (#4319), a vanished session file no longer stops the watcher (#3642), and match rules gain all, any, starts_with and not_starts_with (#4223).
• Cloud sync of large prompts. Prompt text is cut to 200,000 bytes before it is read, which ended SQLiteError: out of memory crashes, and a prompt too large for one sync body uploads truncated with a marker (#3537). Prompts dead-lettered for size before this are re-queued once (#4311).
• Cloud sync pull and status. One pulled operation that can never apply is set aside in sync_pull_quarantine instead of stopping the pull, and /api/sync/status reports pullQuarantine (#4346). Status probes share one request with a 20-second limit, and background pulls wait out the retry delay after a failure (#4330).

Security and privacy

• Opt-in secret redaction. With CLAUDE_MEM_REDACT_ENABLED=true (default false), 11 built-in patterns (AWS, GitHub, OpenAI, Anthropic, Slack, Stripe and Google keys, JWTs, PEM private keys and claude-mem's own keys) are replaced with markers before tool data, prompts and the last assistant message are stored. CLAUDE_MEM_REDACT_DISABLED_BUILTINS and CLAUDE_MEM_REDACT_CUSTOM_PATTERNS adjust the set, and server-runtime event bodies and tool log lines are covered too. (#2616, #4298)
• Sign-in sends a usage summary. The installer's cmem.ai sign-in request now also carries install_state (fresh or update) and a summary of the last 28 UTC days of local memory: per-day observation counts and discovery-token sums only, with no prompts, observation text, paths or project names. The installer prints the same figures as "Your memory so far". (#4266)
• cmem account keys stay on the gateway. A cm_pro_ key is sent only to the cmem.ai gateway, from every key pool and the server runtime; before, one placed in a Gemini key setting was sent to Google (#4309, #4276). GET /api/settings also masks key pools stored as JSON arrays (#4309).
• settings.json is owner-only. Every writer creates ~/.claude-mem/settings.json with mode 0600, and the cloud-sync skill no longer asks you to paste a sync token into chat. (#3498)
• Excluded projects stay excluded. Summaries skip unknown sessions and excluded checkouts, so an excluded OpenCode checkout no longer gets a session and an observer call every turn (#4310), and UserPromptSubmit honors CLAUDE_MEM_EXCLUDED_PROJECTS (#3311).
• Less text leaves the machine. Telegram session wrap-ups give file counts instead of paths (#4145), and recalled titles written into Grok Bot memory are stripped of control characters and fenced as untrusted text (#4146).
• Telemetry. New worker_start_failed and supervisor_registry_degraded events, a top_abort_reason field, rate_limit, auth, deadline_exceeded, output_retry and drift in abort_reason, and chroma_zero_results in fallback_reason. Telemetry stays opt-out with npx claude-mem telemetry disable.

Windows

• Worker starts. Workers start through Start-Process -WindowStyle Hidden instead of Node's detached spawn, so no console window opens (#3529). The 2-minute spawn cooldown now starts only when a worker crashed during boot or could not be launched, not after any failed start (#3408).
• Ports and locks. The worker's listening socket is no longer inherited by child processes, a spawn.lock held by a dead PID is broken (#3309), and a start-token probe that times out no longer removes the PID-reuse guard (#4165).
• Launching tools. PATH lookups run where.exe directly (#3321), and an observer executable that cannot be spawned (a missing binary, or a .cmd shim Node refuses) is reported once as a setup error, with native .exe files preferred (#4122).
• Dates. When Bun cannot resolve the system time zone, dates fall back to a fixed UTC or ISO form instead of dropping SessionStart context or failing search. (#4126, #4302)

Full Changelog: v13.28.0...v13.29.0