om-memory-system 6.1.0-next.237 → 6.1.1

package/dist/config-template.d.ts

@@ -3,4 +3,4 @@
  * src/config.ts until that file reached its size ceiling; the text is
  * unchanged. Shared by config.ts and the global settings writer.
  */
-export declare const CONFIG_TEMPLATE = "{\n  // ============================================\n  // omms (Opinionated Modular Memory System) Configuration\n  // ============================================\n  \n  // Storage location for the vector database. Leave unset to use the\n  // default ~/.omms/data (a legacy ~/.opencode-mem/data store is migrated\n  // there automatically on first start; see docs/omms-migration.md).\n  // \"storagePath\": \"~/.omms/data\",\n\n  \"userEmailOverride\": \"\",\n  \"userNameOverride\": \"\",\n  \n  // ============================================\n  // Embedding Model (for similarity search)\n  // ============================================\n  // Local = Hugging Face / ONNX via @huggingface/transformers (not Apple MLX).\n  // Remote = set embeddingApiUrl (OpenAI-compatible /embeddings); embeddingApiKey is optional.\n  \n  // Default: Nomic Embed v1 (768 dimensions, 8192 context, multilingual)\n  \"embeddingModel\": \"Xenova/nomic-embed-text-v1\",\n\n  // Opt-in Nomic task prefixes (search_document: / search_query:). After enabling,\n  // re-index existing memories so store and query vectors stay aligned.\n  // \"embeddingUseTaskPrefixes\": true,\n  \n  // Auto-detected dimensions (no need to set manually)\n  // \"embeddingDimensions\": 768,\n  \n  // Other recommended local models:\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-base-en\",  // 768 dims, English-only, 8192 context\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-small-en\", // 512 dims, faster, 8192 context\n  // \"embeddingModel\": \"Xenova/all-MiniLM-L6-v2\",            // 384 dims, very fast, 512 context\n  // \"embeddingModel\": \"Xenova/all-mpnet-base-v2\",           // 768 dims, good quality, 512 context\n  \n  // Optional: OpenAI-compatible API for embeddings (URL required; key only if the server needs one)\n  // \"embeddingApiUrl\": \"https://api.openai.com/v1\",\n  // \"embeddingApiKey\": \"env://OPENAI_API_KEY\",  // or \"sk-...\" / \"file:///path/to/key\"\n  // \"embeddingModel\": \"text-embedding-3-small\",  // 1536 dims, auto-detected\n \n  // ============================================\n  // Web Server Settings\n  // ============================================\n  \n  // Start a background import of past chats on each host's next start.\n  \"autoBackfill\": true,\n  // \"piBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // \"opencodeBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // Directory maps for history recorded in moved or deleted directories.\n  // \"importPathMaps\": [{ \"from\": \"~/code/app-feat-x\", \"to\": \"~/code/app\" }],\n  // Unresolved directories to hide on the Directory maps page. They stay unimported.\n  // \"importIgnoredDirectories\": [\"/private/tmp/scratch\"],\n  // Claude Code's folder, when it is not ~/.claude (like CLAUDE_CONFIG_DIR).\n  // \"claudeConfigDir\": \"~/.claude-work\",\n\n  // Register the web app to start when you log in.\n  \"webServerAutoStart\": true,\n  // Enable web UI for managing memories (accessible at http://localhost:4747)\n  \"webServerEnabled\": true,\n  \n  // Port for web UI server\n  \"webServerPort\": 4747,\n  \n  // Host address for web UI (use 127.0.0.1 for local only, 0.0.0.0 for network access)\n  \"webServerHost\": \"127.0.0.1\",\n\n  // Optional HTTP Basic Auth for the web UI. Accepts literal, env://, or file:// secrets.\n  // \"webServerAuthPassword\": \"\",\n  // \"webServerAuthUsername\": \"\",\n\n  // Deprecated: imported once into the API token table, then no longer read.\n  // Create API tokens on the Settings page (Keys and access) instead.\n  // \"webServerApiToken\": \"env://OMMS_WEB_TOKEN\",\n  \n  // ============================================\n  // Database Settings\n  // ============================================\n  \n  // Maximum vectors per database shard (auto-creates new shard when limit reached)\n  \"maxVectorsPerShard\": 50000,\n  \n  // Automatically delete old memories based on retention period\n  \"autoCleanupEnabled\": true,\n  \n  // Days to keep memories before auto-cleanup (only if autoCleanupEnabled is true)\n  \"autoCleanupRetentionDays\": 30,\n  \n  // Automatically detect and remove duplicate memories\n  \"deduplicationEnabled\": true,\n  \n   // Similarity threshold (0-1) for detecting duplicates (higher = stricter)\n   \"deduplicationSimilarityThreshold\": 0.90,\n   \n  // ============================================\n  // Memory Scope Settings\n  // ============================================\n\n  // Default scope for memory list/search queries\n  // \"project\" keeps queries within the current project, \"all-projects\" searches across all project shards\n  \"memory\": {\n    \"defaultScope\": \"project\"\n  },\n\n  // ============================================\n  // Model for auto-capture and profile learning (same rule in OpenCode and Pi)\n  // ============================================\n\n   // Which model summarises your work, in this order:\n   //   1. The host model below: opencodeProvider/opencodeModel in OpenCode,\n   //      piProvider/piModel in Pi. Set the model to \"inherit\" to follow\n   //      whatever model the session is using, or to \"external\" to send\n   //      every call to the external API below.\n   //   2. If no host model is set: the external API (memoryModel/memoryApiUrl/\n   //      memoryApiKey further down).\n   //   3. If neither is set: the session's own model.\n   // If the host model fails and the external API is configured, the external\n   // API is used instead.\n   //\n   // Host models go through OpenCode's or Pi's own sign-in (OAuth like Claude\n   // Pro/Max, GitHub Copilot, ChatGPT, or your own key there), so no separate API\n   // key is needed here.\n   //\n   // Examples (OpenCode names from 'opencode providers list'; Pi names from its model list):\n   //   \"opencodeProvider\": \"anthropic\",    \"opencodeModel\": \"claude-haiku-4-5-20251001\"\n   //   \"opencodeProvider\": \"openai\",       \"opencodeModel\": \"inherit\"\n   //   \"piProvider\": \"openai-codex\",       \"piModel\": \"gpt-5.6-luna\"\n   //\n   // \"opencodeProvider\": \"anthropic\",\n   // \"opencodeModel\": \"claude-haiku-4-5-20251001\",\n   // \"piProvider\": \"openai-codex\",\n   // \"piModel\": \"gpt-5.6-luna\",\n\n   // ============================================\n   // Auto-Capture Settings\n   // ============================================\n  \n  // Auto-capture runs in the background without blocking your main session.\n  // Note: Ollama may not support tool calling. Use OpenAI, Anthropic, or Groq for best results.\n  \n  \"autoCaptureEnabled\": true,\n  \n  // Provider type: \"openai-chat\" | \"openai-responses\" | \"anthropic\" | \"minimax\" | \"orcarouter\" | \"google-gemini\"\n  // Note: \"openai-chat\" is a generic OpenAI API-compatible mode.\n  // Any service that follows the OpenAI Chat Completions API can use it via custom \"memoryApiUrl\".\n  \"memoryProvider\": \"openai-chat\",\n  \n  // External API (step 2 above, and the fallback when a host model fails).\n  // Uncomment all 3 lines and replace memoryApiKey before use:\n  // \"memoryModel\": \"gpt-4o-mini\",\n  // \"memoryApiUrl\": \"https://api.openai.com/v1\",\n  // \"memoryApiKey\": \"sk-...\",\n\n  // API Key Formats:\n  // Direct value:        \"sk-...\"\n  // From file:           \"file://~/.config/litellm-key.txt\"\n  // From env variable:   \"env://LITELLM_API_KEY\"\n  \n  // Examples for different providers:\n  // Any OpenAI-compatible endpoint can use the \"openai-chat\" provider pattern below.\n  // Common examples: DeepSeek, Qwen (via Alibaba Cloud ModelStudio),\n  // Zhipu GLM (BigModel platform), and Kimi (Moonshot AI platform).\n\n  // OpenAI Chat Completion (default, backward compatible):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"gpt-4o-mini\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n\n  // DeepSeek (OpenAI-compatible example):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"deepseek-chat\"\n  //   \"memoryApiUrl\": \"https://api.deepseek.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // OpenAI Responses API (recommended, with session support):\n  //   \"memoryProvider\": \"openai-responses\"\n  //   \"memoryModel\": \"gpt-4o\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // Anthropic (with session support):\n  //   \"memoryProvider\": \"anthropic\"\n  //   \"memoryModel\": \"claude-3-5-haiku-20241022\"\n  //   \"memoryApiUrl\": \"https://api.anthropic.com/v1\"\n  //   \"memoryApiKey\": \"sk-ant-...\"\n\n  // MiniMax (Anthropic Messages-compatible endpoint, with session support):\n  //   \"memoryProvider\": \"minimax\"\n  //   \"memoryModel\": \"MiniMax-M3\"\n  //   \"memoryApiUrl\": \"https://api.minimax.io\"        // global endpoint\n  //   \"memoryApiKey\": \"<MiniMax API key>\"\n  //   // China endpoint: \"memoryApiUrl\": \"https://api.minimaxi.com\"\n  //   // Optional adaptive thinking for MiniMax-M3:\n  //   \"memoryExtraParams\": { \"thinking\": { \"type\": \"adaptive\" } }\n\n  // OrcaRouter (OpenAI-compatible gateway, namespaced model IDs, with session support):\n  //   \"memoryProvider\": \"orcarouter\"\n  //   \"memoryApiKey\": \"<OrcaRouter API key>\"\n  //   // memoryApiUrl and memoryModel are optional \u2014 they default to\n  //   // https://api.orcarouter.ai/v1 and \"orcarouter/auto\" (a routing alias).\n  //   // OrcaRouter rejects bare model names, so if you set memoryModel, use a\n  //   // namespaced ID such as \"openai/gpt-5.5\" or \"deepseek/deepseek-v4-flash\".\n  //   \"memoryModel\": \"openai/gpt-5.5\"\n\n  // Groq (OpenAI-compatible, use openai-chat provider):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"llama-3.3-70b-versatile\"\n  //   \"memoryApiUrl\": \"https://api.groq.com/openai/v1\"\n  //   \"memoryApiKey\": \"gsk_...\"\n  \n  // Maximum iterations for multi-turn AI analysis (for openai-responses, anthropic, and minimax)\n  \"autoCaptureMaxIterations\": 5,\n   \n  // Timeout per iteration in milliseconds (30 seconds default)\n  \"autoCaptureIterationTimeout\": 30000,\n\n  // Tries for one turn's capture, on Pi and OpenCode, 2 and 4 seconds apart. After the\n  // last one fails, a turn the model could not reach goes to the capture retry queue.\n  \"autoCaptureMaxRetries\": 3,\n\n  // Maximum UTF-8 bytes for the auto-capture markdown context sent to the summary model.\n  // Prevents HTTP 400 context overflows on models with ~131K token windows (e.g. Groq Llama).\n  // Rough guide: tokens \u2248 bytes / 4 for mixed code/prose.\n  \"autoCaptureMaxContextBytes\": 131072,\n   \n  // Days to keep AI session history before cleanup\n  \"aiSessionRetentionDays\": 7,\n\n  // Capture diagnostics: every capture attempt always writes one metadata line\n  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no\n  // conversation text. Set captureTrace to true to also write each attempt's\n  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain\n  // conversation content: <private> text and common API key formats are\n  // redacted, files are readable only by you, and they are deleted after\n  // captureTraceRetentionDays. Only this global file can turn tracing on.\n  // \"captureTrace\": false,\n  // \"captureTraceRetentionDays\": 7,\n  // \"captureAttemptRetentionDays\": 30,\n\n  // Skill learning: record each full turn (prompt, replies, tool calls with\n  // their outcome, and the skills used) in ~/.omms/data/skills.db, so OMMS can\n  // rank your skills and draft better versions. Off by default. A project config\n  // can turn it off but never on. No model call is made to record a turn.\n  // \"skillLearning\": false,\n  // Remove <private> text and common API key formats before a turn is stored.\n  // Global config only. Turning this off stores text as written.\n  // \"skillLearningRedact\": true,\n  // Folders OMMS reads skills from. Absolute or starting with ~/. Global config\n  // only. Unset: ~/.agents/skills and the Claude Code folder's skills folder.\n  // \"skillFolders\": [\"~/.agents/skills\"],\n\n  // Visible recall: the OpenCode terminal TUI and Pi's interactive terminal draw\n  // a row for each memory recall. On by default. A project config can turn it\n  // off but never on.\n  // \"visibleRecall\": true,\n\n  // Memory injection, on OpenCode V2, Pi and Claude Code. Each prompt gets the\n  // memories that match it. openingContext also searches memory with\n  // openingContextQuery and adds the matches to the\n  // first message of a new session. The query is global only, trimmed, and\n  // 500 characters at most. subagentMemory gives each subagent the memories\n  // that match its task. For both switches, a project config can only turn these off.\n  // \"openingContext\": false,\n  // \"openingContextQuery\": \"project architecture, setup, conventions and key decisions\",\n  // \"subagentMemory\": true,\n\n  // Capture retry queue: when a live capture fails because the model cannot be\n  // reached (network error, timeout, HTTP 408, 429 or 5xx), OMMS keeps a copy of\n  // the turn in ~/.omms/data and tries it again later. The copy has <private>\n  // text and common API key formats removed. Turns are deleted when the retry\n  // ends or after captureRetryRetentionHours (0 to 720). 0 turns the queue off.\n  // Only this global file can set it.\n  // \"captureRetryRetentionHours\": 72,\n\n  // Temperature for AI API requests (set to false to omit parameter for models that don't support it)\n  // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter\n  // Set to false and add \"memoryTemperature\": false in config when using such models\n  \"memoryTemperature\": 0.3,\n\n  // Extra parameters to include in API request body\n  // Useful for local inference servers (e.g. llama-server with --jinja) that support\n  // additional parameters like disabling thinking/reasoning mode\n  // Example for Qwen3 models: { \"enable_thinking\": false }\n  // \"memoryExtraParams\": {},\n\n  // Language for auto-capture summaries (default: \"auto\" for auto-detection)\n  // Options: \"auto\", \"en\", \"id\", \"zh\", \"ja\", \"es\", \"fr\", \"de\", \"ru\", \"pt\", \"ar\", \"ko\"\n  // \"autoCaptureLanguage\": \"auto\",\n\n  // ============================================\n  // Toast Notifications\n  // ============================================\n\n  // Show toast when memory is auto-captured\n  \"showAutoCaptureToasts\": true,\n\n  // Show toast when user profile is updated\n  \"showUserProfileToasts\": true,\n\n  // Show toast for error messages\n  \"showErrorToasts\": true,\n\n  // ============================================\n  // User Profile System\n  // ============================================\n\n  // Analyze user prompts every N prompts to build/update your user profile\n  // When N uncaptured prompts accumulate, AI will analyze them to identify:\n  // - User preferences (code style, communication style, tool preferences)\n  // - User patterns (recurring topics, problem domains, technical interests)\n  // - User workflows (development habits, sequences, learning style)\n  // - Skill level (overall and per-domain assessment)\n  \"userProfileAnalysisInterval\": 10,\n\n  // Days before inactive items (all types) are eligible for removal\n  \"userProfileStaleDays\": 2,\n\n  // Days before an inactive workflow is eligible for removal. A workflow seen\n  // userProfileMinEvidenceForRetention times or more is kept.\n  \"userProfileWorkflowStaleDays\": 30,\n\n  // Number of preferences shown in UI\n  \"userProfileDisplayPreferences\": 20,\n  \n  // Number of patterns shown in UI\n  \"userProfileDisplayPatterns\": 15,\n  \n  // Number of workflows shown in UI\n  \"userProfileDisplayWorkflows\": 10,\n  \n  // Number of preferences injected into LLM conversation context\n  // Keep this small \u2014 the strongest signals are enough; more dilute LLM attention\n  \"userProfileInjectPreferences\": 5,\n  \n  // Number of patterns injected into LLM conversation context\n  \"userProfileInjectPatterns\": 5,\n  \n  // Number of workflows injected into LLM conversation context\n  \"userProfileInjectWorkflows\": 3,\n  \n  // Days before preference confidence starts to decay (if not reinforced)\n  // Preferences that aren't seen again will gradually lose confidence and be removed\n  \"userProfileConfidenceDecayDays\": 30,\n  \n  // Number of profile versions to keep in changelog (for rollback/debugging)\n  // Older versions are automatically cleaned up\n  \"userProfileChangelogRetentionCount\": 5,\n\n  // Minimum evidence count for a preference/pattern to survive confidence decay\n  // Items confirmed fewer times are more likely to be pruned when confidence decays\n  \"userProfileMinEvidenceForRetention\": 3,\n\n  // Merge duplicate or vague profile items every N analysed prompts (about 1 model call per 100 items)\n  \"userProfileAutoCleanupEnabled\": true,\n  \"userProfileAutoCleanupInterval\": 500,\n\n  // Enable LLM validation of existing preferences against recent behavior.\n  // When enabled, each analysis round checks if top-5 preferences still match recent prompts.\n  // Experimental \u2014 disabled by default.\n  \"userProfileValidationEnabled\": false,\n\n  // ============================================\n  // Search Settings\n  // ============================================\n   \n  // Minimum similarity score (0-1) for memory search results. Leave unset to\n  // follow the default for your embedding model: 0.5 for qwen3-embedding:0.6b\n  // (the only model measured), 0.6 for every other embedder.\n  // \"similarityThreshold\": 0.5,\n\n  // Maximum number of memories to return in search results.\n  // Manual searches can ask for fewer; prompt retrieval uses this ceiling.\n  \"maxMemories\": 10,\n\n  // ============================================\n  // Memory Context Controls\n  // ============================================\n\n  // Per-prompt memory retrieval on every host. \"enabled\": false turns it off;\n  // \"excludeCurrentSession\" leaves out memories captured in the current session.\n  \"chatMessage\": {\n    \"enabled\": true,\n    \"excludeCurrentSession\": true,\n  },\n\n  // OpenCode profile-learning input limit, in UTF-8 bytes. Smaller values can\n  // omit prompts from that input. Does not limit Pi or Claude Code profile input.\n  \"userProfileMaxContextBytes\": 32768,\n\n  // Shared budget for automatic memory context (profile text, memories, and\n  // formatting) across all hosts, in approximate tokens: ceil(UTF-8 bytes / 4).\n  // Covers prompt retrieval, recent-memory injection, and session-memory\n  // restoration. Not an exact provider token limit or a spending limit.\n  \"retrievalMaxTokens\": 2000,\n\n  // Memorisation settings (global config only).\n  // Repeat for a session snapshot: \"add\" or \"replace\" (default \"add\").\n  // \"memorisationRepeat\": \"add\",\n  // Maximum stored text: 100 to 20,000 characters or 25 to 5,000 tokens.\n  // Tokens are estimated with the same rule as retrievalMaxTokens.\n  // \"memorisationMaxSize\": 2000,\n  // \"memorisationSizeUnit\": \"characters\", // or \"tokens\"\n\n  // Restore a session's own memories after compaction and after resume, on\n  // OpenCode V2, Pi and Claude Code. memoryLimit is the most memories to\n  // restore. A project config can turn it on or off.\n  // \"compaction\": { \"enabled\": true, \"memoryLimit\": 10 },\n\n  // ============================================\n  // Advanced Settings\n  // ============================================\n  \n  // Inject user profile into AI context (preferences, patterns, workflows)\n  \"injectProfile\": true\n}\n";
+export declare const CONFIG_TEMPLATE = "{\n  // ============================================\n  // omms (Opinionated Modular Memory System) Configuration\n  // ============================================\n  \n  // Storage location for the vector database. Leave unset to use the\n  // default ~/.omms/data (a legacy ~/.opencode-mem/data store is migrated\n  // there automatically on first start; see docs/omms-migration.md).\n  // \"storagePath\": \"~/.omms/data\",\n\n  \"userEmailOverride\": \"\",\n  \"userNameOverride\": \"\",\n  \n  // ============================================\n  // Embedding Model (for similarity search)\n  // ============================================\n  // Local = Hugging Face / ONNX via @huggingface/transformers (not Apple MLX).\n  // Remote = set embeddingApiUrl (OpenAI-compatible /embeddings); embeddingApiKey is optional.\n  \n  // Default: Nomic Embed v1 (768 dimensions, 8192 context, multilingual)\n  \"embeddingModel\": \"Xenova/nomic-embed-text-v1\",\n\n  // Opt-in Nomic task prefixes (search_document: / search_query:). After enabling,\n  // re-index existing memories so store and query vectors stay aligned.\n  // \"embeddingUseTaskPrefixes\": true,\n  \n  // Auto-detected dimensions (no need to set manually)\n  // \"embeddingDimensions\": 768,\n  \n  // Other recommended local models:\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-base-en\",  // 768 dims, English-only, 8192 context\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-small-en\", // 512 dims, faster, 8192 context\n  // \"embeddingModel\": \"Xenova/all-MiniLM-L6-v2\",            // 384 dims, very fast, 512 context\n  // \"embeddingModel\": \"Xenova/all-mpnet-base-v2\",           // 768 dims, good quality, 512 context\n  \n  // Optional: OpenAI-compatible API for embeddings (URL required; key only if the server needs one)\n  // \"embeddingApiUrl\": \"https://api.openai.com/v1\",\n  // \"embeddingApiKey\": \"env://OPENAI_API_KEY\",  // or \"sk-...\" / \"file:///path/to/key\"\n  // \"embeddingModel\": \"text-embedding-3-small\",  // 1536 dims, auto-detected\n \n  // ============================================\n  // Web Server Settings\n  // ============================================\n  \n  // Start a background import of past chats on each host's next start.\n  \"autoBackfill\": true,\n  // \"piBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // \"opencodeBackfillModel\": \"inherit\", // or \"external\", or \"provider/model\"\n  // Directory maps for history recorded in moved or deleted directories.\n  // \"importPathMaps\": [{ \"from\": \"~/code/app-feat-x\", \"to\": \"~/code/app\" }],\n  // Unresolved directories to hide on the Directory maps page. They stay unimported.\n  // \"importIgnoredDirectories\": [\"/private/tmp/scratch\"],\n  // Claude Code's folder, when it is not ~/.claude (like CLAUDE_CONFIG_DIR).\n  // \"claudeConfigDir\": \"~/.claude-work\",\n\n  // Register the web app to start when you log in.\n  \"webServerAutoStart\": true,\n  // Enable web UI for managing memories (accessible at http://localhost:4747)\n  \"webServerEnabled\": true,\n  \n  // Port for web UI server\n  \"webServerPort\": 4747,\n  \n  // Host address for web UI (use 127.0.0.1 for local only, 0.0.0.0 for network access)\n  \"webServerHost\": \"127.0.0.1\",\n\n  // Optional HTTP Basic Auth for the web UI. Accepts literal, env://, or file:// secrets.\n  // \"webServerAuthPassword\": \"\",\n  // \"webServerAuthUsername\": \"\",\n\n  // Deprecated: imported once into the API token table, then no longer read.\n  // Create API tokens on the Settings page (Keys and access) instead.\n  // \"webServerApiToken\": \"env://OMMS_WEB_TOKEN\",\n  \n  // ============================================\n  // Database Settings\n  // ============================================\n  \n  // Maximum vectors per database shard (auto-creates new shard when limit reached)\n  \"maxVectorsPerShard\": 50000,\n  \n  // Automatically delete old memories based on retention period\n  \"autoCleanupEnabled\": true,\n  \n  // Days to keep memories before auto-cleanup (only if autoCleanupEnabled is true)\n  \"autoCleanupRetentionDays\": 30,\n  \n  // Automatically detect and remove duplicate memories\n  \"deduplicationEnabled\": true,\n  \n   // Similarity threshold (0-1) for detecting duplicates (higher = stricter)\n   \"deduplicationSimilarityThreshold\": 0.90,\n   \n  // ============================================\n  // Memory Scope Settings\n  // ============================================\n\n  // Default scope for memory list/search queries\n  // \"project\" keeps queries within the current project, \"all-projects\" searches across all project shards\n  \"memory\": {\n    \"defaultScope\": \"project\"\n  },\n\n  // ============================================\n  // Model for auto-capture and profile learning (same rule in OpenCode and Pi)\n  // ============================================\n\n   // Which model summarises your work, in this order:\n   //   1. The host model below: opencodeProvider/opencodeModel in OpenCode,\n   //      piProvider/piModel in Pi. Set the model to \"inherit\" to follow\n   //      whatever model the session is using, or to \"external\" to send\n   //      every call to the external API below.\n   //   2. If no host model is set: the external API (memoryModel/memoryApiUrl/\n   //      memoryApiKey further down).\n   //   3. If neither is set: the session's own model.\n   // If the host model fails and the external API is configured, the external\n   // API is used instead.\n   //\n   // Host models go through OpenCode's or Pi's own sign-in (OAuth like Claude\n   // Pro/Max, GitHub Copilot, ChatGPT, or your own key there), so no separate API\n   // key is needed here.\n   //\n   // Examples (OpenCode names from 'opencode providers list'; Pi names from its model list):\n   //   \"opencodeProvider\": \"anthropic\",    \"opencodeModel\": \"claude-haiku-4-5-20251001\"\n   //   \"opencodeProvider\": \"openai\",       \"opencodeModel\": \"inherit\"\n   //   \"piProvider\": \"openai-codex\",       \"piModel\": \"gpt-5.6-luna\"\n   //\n   // \"opencodeProvider\": \"anthropic\",\n   // \"opencodeModel\": \"claude-haiku-4-5-20251001\",\n   // \"piProvider\": \"openai-codex\",\n   // \"piModel\": \"gpt-5.6-luna\",\n\n   // ============================================\n   // Auto-Capture Settings\n   // ============================================\n  \n  // Auto-capture runs in the background without blocking your main session.\n  // Note: Ollama may not support tool calling. Use OpenAI, Anthropic, or Groq for best results.\n  \n  \"autoCaptureEnabled\": true,\n  \n  // Provider type: \"openai-chat\" | \"openai-responses\" | \"anthropic\" | \"minimax\" | \"orcarouter\" | \"google-gemini\"\n  // Note: \"openai-chat\" is a generic OpenAI API-compatible mode.\n  // Any service that follows the OpenAI Chat Completions API can use it via custom \"memoryApiUrl\".\n  \"memoryProvider\": \"openai-chat\",\n  \n  // External API (step 2 above, and the fallback when a host model fails).\n  // Uncomment all 3 lines and replace memoryApiKey before use:\n  // \"memoryModel\": \"gpt-4o-mini\",\n  // \"memoryApiUrl\": \"https://api.openai.com/v1\",\n  // \"memoryApiKey\": \"sk-...\",\n\n  // API Key Formats:\n  // Direct value:        \"sk-...\"\n  // From file:           \"file://~/.config/litellm-key.txt\"\n  // From env variable:   \"env://LITELLM_API_KEY\"\n  \n  // Examples for different providers:\n  // Any OpenAI-compatible endpoint can use the \"openai-chat\" provider pattern below.\n  // Common examples: DeepSeek, Qwen (via Alibaba Cloud ModelStudio),\n  // Zhipu GLM (BigModel platform), and Kimi (Moonshot AI platform).\n\n  // OpenAI Chat Completion (default, backward compatible):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"gpt-4o-mini\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n\n  // DeepSeek (OpenAI-compatible example):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"deepseek-chat\"\n  //   \"memoryApiUrl\": \"https://api.deepseek.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // OpenAI Responses API (recommended, with session support):\n  //   \"memoryProvider\": \"openai-responses\"\n  //   \"memoryModel\": \"gpt-4o\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // Anthropic (with session support):\n  //   \"memoryProvider\": \"anthropic\"\n  //   \"memoryModel\": \"claude-3-5-haiku-20241022\"\n  //   \"memoryApiUrl\": \"https://api.anthropic.com/v1\"\n  //   \"memoryApiKey\": \"sk-ant-...\"\n\n  // MiniMax (Anthropic Messages-compatible endpoint, with session support):\n  //   \"memoryProvider\": \"minimax\"\n  //   \"memoryModel\": \"MiniMax-M3\"\n  //   \"memoryApiUrl\": \"https://api.minimax.io\"        // global endpoint\n  //   \"memoryApiKey\": \"<MiniMax API key>\"\n  //   // China endpoint: \"memoryApiUrl\": \"https://api.minimaxi.com\"\n  //   // Optional adaptive thinking for MiniMax-M3:\n  //   \"memoryExtraParams\": { \"thinking\": { \"type\": \"adaptive\" } }\n\n  // OrcaRouter (OpenAI-compatible gateway, namespaced model IDs, with session support):\n  //   \"memoryProvider\": \"orcarouter\"\n  //   \"memoryApiKey\": \"<OrcaRouter API key>\"\n  //   // memoryApiUrl and memoryModel are optional \u2014 they default to\n  //   // https://api.orcarouter.ai/v1 and \"orcarouter/auto\" (a routing alias).\n  //   // OrcaRouter rejects bare model names, so if you set memoryModel, use a\n  //   // namespaced ID such as \"openai/gpt-5.5\" or \"deepseek/deepseek-v4-flash\".\n  //   \"memoryModel\": \"openai/gpt-5.5\"\n\n  // Groq (OpenAI-compatible, use openai-chat provider):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"llama-3.3-70b-versatile\"\n  //   \"memoryApiUrl\": \"https://api.groq.com/openai/v1\"\n  //   \"memoryApiKey\": \"gsk_...\"\n  \n  // Maximum iterations for multi-turn AI analysis (for openai-responses, anthropic, and minimax)\n  \"autoCaptureMaxIterations\": 5,\n   \n  // Timeout per iteration in milliseconds (30 seconds default)\n  \"autoCaptureIterationTimeout\": 30000,\n\n  // Tries for one turn's capture, on Pi and OpenCode, 2 and 4 seconds apart. After the\n  // last one fails, a turn the model could not reach goes to the capture retry queue.\n  \"autoCaptureMaxRetries\": 3,\n\n  // Maximum UTF-8 bytes for the auto-capture markdown context sent to the summary model.\n  // Prevents HTTP 400 context overflows on models with ~131K token windows (e.g. Groq Llama).\n  // Rough guide: tokens \u2248 bytes / 4 for mixed code/prose.\n  \"autoCaptureMaxContextBytes\": 131072,\n   \n  // Days to keep AI session history before cleanup\n  \"aiSessionRetentionDays\": 7,\n\n  // Capture diagnostics: every capture attempt always writes one metadata line\n  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no\n  // conversation text. Set captureTrace to true to also write each attempt's\n  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain\n  // conversation content: <private> text and common API key formats are\n  // redacted, files are readable only by you, and they are deleted after\n  // captureTraceRetentionDays. Only this global file can turn tracing on.\n  // \"captureTrace\": false,\n  // \"captureTraceRetentionDays\": 7,\n  // \"captureAttemptRetentionDays\": 30,\n\n  // Skill learning: record each full turn (prompt, replies, tool calls with\n  // their outcome, and the skills used) in ~/.omms/data/skills.db, so OMMS can\n  // rank your skills and draft better versions. Off by default. A project config\n  // can turn it off but never on. No model call is made to record a turn.\n  // \"skillLearning\": false,\n  // Remove <private> text and common API key formats before a turn is stored.\n  // Global config only. Turning this off stores text as written.\n  // \"skillLearningRedact\": true,\n  // Folders OMMS reads skills from. Absolute or starting with ~/. Global config\n  // only. Unset: ~/.agents/skills and the Claude Code folder's skills folder.\n  // \"skillFolders\": [\"~/.agents/skills\"],\n\n  // Visible recall: the OpenCode terminal TUI and Pi's interactive terminal draw\n  // a row for each memory recall. On by default. A project config can turn it\n  // off but never on.\n  // \"visibleRecall\": true,\n\n  // Memory injection, on OpenCode V2, Pi and Claude Code. Each prompt gets the\n  // memories that match it. openingContext also searches memory with\n  // openingContextQuery and adds the matches to the\n  // first message of a new session. The query is global only, trimmed, and\n  // 500 characters at most. subagentMemory gives each subagent the memories\n  // that match its task. For both switches, a project config can only turn these off.\n  // \"openingContext\": false,\n  // \"openingContextQuery\": \"project architecture, setup, conventions and key decisions\",\n  // \"subagentMemory\": true,\n\n  // Capture retry queue: when a live capture fails because the model cannot be\n  // reached (network error, timeout, HTTP 408, 429 or 5xx), OMMS keeps a copy of\n  // the turn in ~/.omms/data and tries it again later. The copy has <private>\n  // text and common API key formats removed. Turns are deleted when the retry\n  // ends or after captureRetryRetentionHours (0 to 720). 0 turns the queue off.\n  // Only this global file can set it.\n  // \"captureRetryRetentionHours\": 72,\n\n  // Temperature for AI API requests (set to false to omit parameter for models that don't support it)\n  // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter\n  // Set to false and add \"memoryTemperature\": false in config when using such models\n  \"memoryTemperature\": 0.3,\n\n  // Extra parameters to include in API request body\n  // Useful for local inference servers (e.g. llama-server with --jinja) that support\n  // additional parameters like disabling thinking/reasoning mode\n  // Example for Qwen3 models: { \"enable_thinking\": false }\n  // \"memoryExtraParams\": {},\n\n  // Language for auto-capture summaries (default: \"auto\" for auto-detection)\n  // Options: \"auto\", \"en\", \"id\", \"zh\", \"ja\", \"es\", \"fr\", \"de\", \"ru\", \"pt\", \"ar\", \"ko\"\n  // \"autoCaptureLanguage\": \"auto\",\n\n  // ============================================\n  // Toast Notifications\n  // ============================================\n\n  // Show toast when memory is auto-captured\n  \"showAutoCaptureToasts\": true,\n\n  // Show toast when user profile is updated\n  \"showUserProfileToasts\": true,\n\n  // Show toast for error messages\n  \"showErrorToasts\": true,\n\n  // ============================================\n  // User Profile System\n  // ============================================\n\n  // Analyze user prompts every N prompts to build/update your user profile\n  // When N uncaptured prompts accumulate, AI will analyze them to identify:\n  // - User preferences (code style, communication style, tool preferences)\n  // - User patterns (recurring topics, problem domains, technical interests)\n  // - User workflows (development habits, sequences, learning style)\n  // - Skill level (overall and per-domain assessment)\n  \"userProfileAnalysisInterval\": 10,\n\n  // Days before inactive items (all types) are eligible for removal\n  \"userProfileStaleDays\": 2,\n\n  // Days before an inactive workflow is eligible for removal. A workflow seen\n  // userProfileMinEvidenceForRetention times or more is kept.\n  \"userProfileWorkflowStaleDays\": 30,\n\n  // Number of preferences shown in UI\n  \"userProfileDisplayPreferences\": 20,\n  \n  // Number of patterns shown in UI\n  \"userProfileDisplayPatterns\": 15,\n  \n  // Number of workflows shown in UI\n  \"userProfileDisplayWorkflows\": 10,\n  \n  // Number of preferences injected into LLM conversation context\n  // Keep this small \u2014 the strongest signals are enough; more dilute LLM attention\n  \"userProfileInjectPreferences\": 5,\n  \n  // Number of patterns injected into LLM conversation context\n  \"userProfileInjectPatterns\": 5,\n  \n  // Number of workflows injected into LLM conversation context\n  \"userProfileInjectWorkflows\": 3,\n  \n  // Days before preference confidence starts to decay (if not reinforced)\n  // Preferences that aren't seen again will gradually lose confidence and be removed\n  \"userProfileConfidenceDecayDays\": 30,\n  \n  // Number of profile versions to keep in changelog (for rollback/debugging)\n  // Older versions are automatically cleaned up\n  \"userProfileChangelogRetentionCount\": 5,\n\n  // Minimum evidence count for a preference/pattern to survive confidence decay\n  // Items confirmed fewer times are more likely to be pruned when confidence decays\n  \"userProfileMinEvidenceForRetention\": 3,\n\n  // Merge duplicate or vague profile items every N analysed prompts (about 1 model call per 100 items)\n  \"userProfileAutoCleanupEnabled\": true,\n  \"userProfileAutoCleanupInterval\": 500,\n\n  // Enable LLM validation of existing preferences against recent behavior.\n  // When enabled, each analysis round checks if top-5 preferences still match recent prompts.\n  // Experimental \u2014 disabled by default.\n  \"userProfileValidationEnabled\": false,\n\n  // ============================================\n  // Search Settings\n  // ============================================\n   \n  // Minimum similarity score (0-1) for memory search results. Leave unset to\n  // follow the default for your embedding model: 0.5 for qwen3-embedding:0.6b\n  // (the only model measured), 0.6 for every other embedder.\n  // \"similarityThreshold\": 0.5,\n\n  // Maximum number of memories to return in search results.\n  // Manual searches can ask for fewer; prompt retrieval uses this ceiling.\n  \"maxMemories\": 10,\n\n  // ============================================\n  // Memory Context Controls\n  // ============================================\n\n  // Per-prompt memory retrieval on every host. \"enabled\": false turns it off;\n  // \"excludeCurrentSession\" leaves out memories captured in the current session.\n  \"chatMessage\": {\n    \"enabled\": true,\n    \"excludeCurrentSession\": true,\n  },\n\n  // OpenCode profile-learning input limit, in UTF-8 bytes. Smaller values can\n  // omit prompts from that input. Does not limit Pi or Claude Code profile input.\n  \"userProfileMaxContextBytes\": 32768,\n\n  // Shared budget for automatic memory context (profile text, memories, and\n  // formatting) across all hosts, in approximate tokens: ceil(UTF-8 bytes / 4).\n  // Covers prompt retrieval, recent-memory injection, and session-memory\n  // restoration. Not an exact provider token limit or a spending limit.\n  \"retrievalMaxTokens\": 2000,\n\n  // Memorisation settings (global config only).\n  // Repeat for a session snapshot: \"add\" or \"replace\" (default \"add\").\n  // \"memorisationRepeat\": \"add\",\n  // Maximum stored text: 100 to 20,000 characters or 25 to 5,000 tokens.\n  // Tokens are estimated with the same rule as retrievalMaxTokens.\n  // \"memorisationMaxSize\": 2000,\n  // \"memorisationSizeUnit\": \"characters\", // or \"tokens\"\n  // Experimental: rerank text recall with the reranker server of the Memory\n  // settings (needs its URL and model; default false).\n  // \"memorisationRerank\": false,\n\n  // Restore a session's own memories after compaction and after resume, on\n  // OpenCode V2, Pi and Claude Code. memoryLimit is the most memories to\n  // restore. A project config can turn it on or off.\n  // \"compaction\": { \"enabled\": true, \"memoryLimit\": 10 },\n\n  // ============================================\n  // Advanced Settings\n  // ============================================\n  \n  // Inject user profile into AI context (preferences, patterns, workflows)\n  \"injectProfile\": true\n}\n";

package/dist/config-template.js

@@ -405,6 +405,9 @@ export const CONFIG_TEMPLATE = `{
   // Tokens are estimated with the same rule as retrievalMaxTokens.
   // "memorisationMaxSize": 2000,
   // "memorisationSizeUnit": "characters", // or "tokens"
+  // Experimental: rerank text recall with the reranker server of the Memory
+  // settings (needs its URL and model; default false).
+  // "memorisationRerank": false,
 
   // Restore a session's own memories after compaction and after resume, on
   // OpenCode V2, Pi and Claude Code. memoryLimit is the most memories to

package/dist/config.d.ts

@@ -33,6 +33,8 @@ interface OmmsConfig {
     memorisationMaxSize?: number;
     memorisationSizeUnit?: memorisation.MemorisationSettings["memorisationSizeUnit"];
     memorisationRepeat?: memorisation.MemorisationSettings["memorisationRepeat"];
+    /** Experimental: rerank text recall with the reranker server. Global config only. */
+    memorisationRerank?: boolean;
     maxProfileItems?: number;
     injectProfile?: boolean;
     containerTagPrefix?: string;
@@ -223,6 +225,7 @@ declare function buildConfig(fileConfig: OmmsConfig): {
     memorisationMaxSize: number;
     memorisationSizeUnit: import("./core/memorisation.js").MemorisationSizeUnit;
     memorisationRepeat: import("./core/memorisation.js").MemorisationRepeat;
+    memorisationRerank: boolean;
     storagePath: string;
     /** The global `storagePath` when a project moved `storagePath`; `skills.db` stays there. */
     skillStoragePath: string | undefined;
@@ -337,6 +340,7 @@ export declare let CONFIG: {
     memorisationMaxSize: number;
     memorisationSizeUnit: import("./core/memorisation.js").MemorisationSizeUnit;
     memorisationRepeat: import("./core/memorisation.js").MemorisationRepeat;
+    memorisationRerank: boolean;
     storagePath: string;
     /** The global `storagePath` when a project moved `storagePath`; `skills.db` stays there. */
     skillStoragePath: string | undefined;

package/dist/config.js

@@ -41,6 +41,7 @@ const DEFAULTS = {
     memorisationMaxSize: 2000,
     memorisationSizeUnit: "characters",
     memorisationRepeat: "add",
+    memorisationRerank: false,
     maxProfileItems: 5,
     injectProfile: true,
     // omms since the tag prefix migration: new memories carry the fork's

package/dist/core/memorisation-operations.d.ts

@@ -1,4 +1,3 @@
-import { type MemoryScope } from "../services/client.js";
 import type { LinkedSearchSettings } from "../services/linked-search.js";
 import type { getTags } from "../services/tags.js";
 import type { MemoryHost } from "../types/index.js";
@@ -16,7 +15,6 @@ export interface RecallArgs {
     /** Comma-separated labels. Only memorisations that carry all of them match. */
     tags?: string;
     limit?: number;
-    scope?: MemoryScope;
 }
 interface MemorisationContext {
     host?: MemoryHost;
@@ -27,7 +25,10 @@ export declare function runMemorise(args: MemoriseArgs, context: MemorisationCon
 /**
  * Fetches memorisations by exact ID, or by text across memorisations only. `tags` keeps only
  * memorisations that carry every label. With `tags` and no text it returns the newest ones.
- * An exact ID ignores `tags`.
+ * An exact ID ignores `tags`. Every read covers the whole active space and its linked spaces,
+ * so a memorisation is found from any folder. `rerank` (the `memorisationRerank` setting)
+ * lets the reranker server reorder the text matches; each match keeps its content-only
+ * confidence, and the one-strong-match rule reads the first two in that order.
  */
-export declare function runRecall(args: RecallArgs, tags: ReturnType<typeof getTags>, scope: MemoryScope | undefined, linked: LinkedSearchSettings): Promise<Record<string, unknown>>;
+export declare function runRecall(args: RecallArgs, linked: LinkedSearchSettings, rerank?: boolean): Promise<Record<string, unknown>>;
 export {};

package/dist/core/memorisation-operations.js

@@ -1,7 +1,7 @@
 import { memoryClient } from "../services/client.js";
 import { findSnapshotByHostSession, getMemorisationById, listMemorisationsByTags, memorisationIdTaken, } from "../services/memorisation-store.js";
 import { isFullyPrivate, stripPrivateContent } from "../services/privacy.js";
-import { hasAllTags, isMemorisationId, MEMORISATION_TAG, memorisationId, memorisationSizeError, parseTagList, pickRecallMatches, snapshotText, visibleTags, } from "./memorisation.js";
+import { hasAllTags, isMemorisationId, MEMORISATION_TAG, memorisationId, memorisationSizeError, parseTagList, pickRecallMatches, snapshotText, visibleTags, withProjectLabel, } from "./memorisation.js";
 const SNAPSHOT_FIELDS = ["topic", "done", "doing", "next"];
 const BLOCKED = { success: false, error: "Private content blocked" };
 /** Strips `<private>` text. A field that is fully private is dropped. */
@@ -54,7 +54,9 @@ export async function runMemorise(args, context, tags, settings, linked) {
     const wantsReplace = built.kind === "snapshot" && settings.memorisationRepeat === "replace";
     const earlier = wantsReplace && sessionId ? await findSnapshotByHostSession(sessionId, tags.project.tag) : null;
     const id = earlier?.id ?? (await memorisationId((candidate) => memorisationIdTaken(candidate, linked)));
-    const recordTags = callerTags(args.tags);
+    // The project name is one more label, so a label filter narrows by project. It comes after
+    // the fixed tag and the caller's labels and is never doubled.
+    const recordTags = withProjectLabel(callerTags(args.tags), tags.project.projectName);
     const info = tags.project;
     const result = await memoryClient.addMemory(built.text, info.tag, {
         id,
@@ -96,30 +98,36 @@ const LABEL_SEARCH_POOL = 200;
 /**
  * Fetches memorisations by exact ID, or by text across memorisations only. `tags` keeps only
  * memorisations that carry every label. With `tags` and no text it returns the newest ones.
- * An exact ID ignores `tags`.
+ * An exact ID ignores `tags`. Every read covers the whole active space and its linked spaces,
+ * so a memorisation is found from any folder. `rerank` (the `memorisationRerank` setting)
+ * lets the reranker server reorder the text matches; each match keeps its content-only
+ * confidence, and the one-strong-match rule reads the first two in that order.
  */
-export async function runRecall(args, tags, scope, linked) {
+export async function runRecall(args, linked, rerank = false) {
     const input = (args.query ?? args.memoryId ?? args.content)?.trim();
     const wanted = parseTagList(args.tags);
     if (!input && wanted.length === 0) {
         return { success: false, error: "query or tags required: a memorisation ID, text or labels" };
     }
     if (!input)
-        return recallByLabels(wanted, tags, scope, args.limit);
+        return recallByLabels(wanted, args.limit);
     if (isMemorisationId(input)) {
-        const found = await getMemorisationById(input, tags.project.tag, linked);
+        const found = await getMemorisationById(input, null, linked);
         if (!found)
             return { success: false, error: `No memorisation has the ID ${input}` };
         return { success: true, count: 1, matches: [found] };
     }
-    const result = await memoryClient.searchMemories(input, tags.project.tag, scope, {
+    const result = await memoryClient.searchMemories(input, "", "all-projects", {
         includeLinked: true,
         kind: "memorisation",
+        reranker: rerank,
+        // A reranked recall still answers "no memorisation matched" when nothing clears the threshold.
+        thresholdAfterRerank: true,
         ...(wanted.length > 0 ? { limit: LABEL_SEARCH_POOL } : {}),
     });
     if (!result.success)
         return { success: false, error: result.error };
-    const candidates = (result.results ?? [])
+    const mapped = (result.results ?? [])
         .filter((hit) => hasAllTags(hit.tags, wanted))
         .map((hit) => ({
         id: hit.id,
@@ -128,8 +136,10 @@ export async function runRecall(args, tags, scope, linked) {
         createdAt: toIso(hit.createdAt),
         tags: visibleTags(hit.tags),
         ...(hit.spaceName ? { space: hit.spaceName } : {}),
-    }))
-        .sort((a, b) => b.confidence - a.confidence);
+    }));
+    // With the reranker on, the client already ordered the own-space rows and the sort would
+    // undo that order. On a reranker failure the client falls back to the confidence order.
+    const candidates = rerank ? mapped : mapped.sort((a, b) => b.confidence - a.confidence);
     const matches = pickRecallMatches(candidates, args.limit);
     return {
         success: true,
@@ -140,8 +150,8 @@ export async function runRecall(args, tags, scope, linked) {
     };
 }
 /** The newest memorisations that carry every label. They are not ranked, so they have no confidence. */
-async function recallByLabels(wanted, tags, scope, limit) {
-    const listed = await listMemorisationsByTags(tags.project.tag, wanted, scope ?? "project", Math.max(1, limit ?? LABEL_RECALL_LIMIT));
+async function recallByLabels(wanted, limit) {
+    const listed = await listMemorisationsByTags("", wanted, "all-projects", Math.max(1, limit ?? LABEL_RECALL_LIMIT));
     if (!listed.success)
         return listed;
     const matches = listed.memories.map((memory) => ({

package/dist/core/memorisation-settings.d.ts

@@ -1,11 +1,11 @@
 import { type MemorisationSettings } from "./memorisation.js";
 export type { MemorisationSettings } from "./memorisation.js";
-export declare const MEMORISATION_KEYS: readonly ["memorisationMaxSize", "memorisationSizeUnit", "memorisationRepeat"];
+export declare const MEMORISATION_KEYS: readonly ["memorisationMaxSize", "memorisationSizeUnit", "memorisationRepeat", "memorisationRerank"];
 export declare const MEMORISATION_DEFAULTS: MemorisationSettings;
 /** Global-only memorisation settings, with defaults for an older config file. */
 export declare function settingsFrom(config: Partial<MemorisationSettings>): MemorisationSettings;
 /** The web writer validates the pair, so a unit edit cannot strand an invalid number. */
 export declare function checkMemorisationSettingsEdits(edits: Record<string, unknown>, current: MemorisationSettings): void;
 export { settingsFrom as memorisationSettingsFrom };
-/** The three settings are global-only; a project config cannot replace them. */
+/** The four settings are global-only; a project config cannot replace them. */
 export declare function removeMemorisationProjectOverrides(config: Record<string, unknown>): void;

package/dist/core/memorisation-settings.js

@@ -3,11 +3,13 @@ export const MEMORISATION_KEYS = [
     "memorisationMaxSize",
     "memorisationSizeUnit",
     "memorisationRepeat",
+    "memorisationRerank",
 ];
 export const MEMORISATION_DEFAULTS = {
     memorisationMaxSize: 2000,
     memorisationSizeUnit: "characters",
     memorisationRepeat: "add",
+    memorisationRerank: false,
 };
 /** Global-only memorisation settings, with defaults for an older config file. */
 export function settingsFrom(config) {
@@ -20,7 +22,11 @@ export function settingsFrom(config) {
     if (memorisationRepeat !== "add" && memorisationRepeat !== "replace") {
         throw new Error("Invalid memorisationRepeat config");
     }
-    return { memorisationMaxSize, memorisationSizeUnit, memorisationRepeat };
+    if (config.memorisationRerank !== undefined && typeof config.memorisationRerank !== "boolean") {
+        throw new Error("Invalid memorisationRerank config");
+    }
+    const memorisationRerank = config.memorisationRerank ?? MEMORISATION_DEFAULTS.memorisationRerank;
+    return { memorisationMaxSize, memorisationSizeUnit, memorisationRepeat, memorisationRerank };
 }
 /** The web writer validates the pair, so a unit edit cannot strand an invalid number. */
 export function checkMemorisationSettingsEdits(edits, current) {
@@ -30,6 +36,9 @@ export function checkMemorisationSettingsEdits(edits, current) {
             throw new Error("memorisationRepeat must be add or replace");
         }
     }
+    if (edits.memorisationRerank !== undefined && typeof edits.memorisationRerank !== "boolean") {
+        throw new Error("memorisationRerank must be true or false");
+    }
     const hasSize = Object.hasOwn(edits, "memorisationMaxSize");
     const hasUnit = Object.hasOwn(edits, "memorisationSizeUnit");
     if (!hasSize && !hasUnit)
@@ -41,7 +50,7 @@ export function checkMemorisationSettingsEdits(edits, current) {
     }
 }
 export { settingsFrom as memorisationSettingsFrom };
-/** The three settings are global-only; a project config cannot replace them. */
+/** The four settings are global-only; a project config cannot replace them. */
 export function removeMemorisationProjectOverrides(config) {
     for (const key of MEMORISATION_KEYS)
         delete config[key];

package/dist/core/memorisation.d.ts

@@ -4,6 +4,7 @@ export interface MemorisationSettings {
     memorisationMaxSize: number;
     memorisationSizeUnit: MemorisationSizeUnit;
     memorisationRepeat: MemorisationRepeat;
+    memorisationRerank: boolean;
 }
 export interface MemorisationSnapshot {
     topic?: string;
@@ -41,3 +42,8 @@ export declare function parseTagList(text: string | undefined): string[];
 export declare function visibleTags(tags: string[] | undefined): string[];
 /** True when the record carries every wanted tag. Case and surrounding spaces do not matter. */
 export declare function hasAllTags(recordTags: string[] | undefined, wanted: string[]): boolean;
+/**
+ * The project name as one more label, after the fixed tag and the caller's labels. The root
+ * folder name `/` is not a usable label, so a caller there gets no project label.
+ */
+export declare function withProjectLabel(labels: string[], projectName: string | undefined): string[];

package/dist/core/memorisation.js

@@ -74,3 +74,13 @@ export function hasAllTags(recordTags, wanted) {
     const have = new Set((recordTags ?? []).map(normaliseTag));
     return wanted.every((tag) => have.has(normaliseTag(tag)));
 }
+/**
+ * The project name as one more label, after the fixed tag and the caller's labels. The root
+ * folder name `/` is not a usable label, so a caller there gets no project label.
+ */
+export function withProjectLabel(labels, projectName) {
+    const name = projectName?.trim().toLowerCase();
+    if (!name || name === "/")
+        return labels;
+    return labels.includes(name) ? labels : [...labels, name];
+}

package/dist/core/memory-operations.js

@@ -176,7 +176,7 @@ export async function executeMemoryOperation(args, context) {
             case "memorize":
                 return await runMemorise(args, context, tags, memorisationSettingsFrom(CONFIG), linkedSettingsFrom(CONFIG));
             case "recall":
-                return await runRecall(args, tags, args.scope ?? CONFIG.memory.defaultScope, linkedSettingsFrom(CONFIG));
+                return await runRecall(args, linkedSettingsFrom(CONFIG), CONFIG.memorisationRerank);
             case "search": {
                 if (!args.query)
                     return { success: false, error: "query required" };

package/dist/services/client.js

@@ -5,7 +5,7 @@ import { tursoConnectionManager } from "./turso/connection-manager.js";
 import { ensureTursoReady } from "./turso/ready.js";
 import { formatTagsForEmbedding } from "./turso/vector-utils.js";
 import { extractScopeFromContainerTag, resolveMemoryScope } from "./memory-scope.js";
-import { rerankSettingsFrom } from "./reranker.js";
+import { rerankSettingsFor } from "./reranker.js";
 import { rerankedSearch } from "./retrieval-rerank.js";
 import { findLinkedMemoryOwner, linkedRefusal, linkedSettingsFrom, listLinkedSpaces, mapListedMemory, searchLinkedSpaces, } from "./linked-search.js";
 import { CONFIG } from "../config.js";
@@ -68,12 +68,14 @@ export class LocalMemoryClient {
         // not change this search's limit or reranker target.
         const maxMemories = options.limit ?? CONFIG.maxMemories;
         const keywordBoost = CONFIG.keywordBoost;
-        const reranker = rerankSettingsFrom(CONFIG);
+        const reranker = rerankSettingsFor(CONFIG, options.reranker);
         // The reranker path below always uses the boost: it needs a full pool on the similarity scale.
         const keywordMode = CONFIG.keywordMode;
-        if (CONFIG.rerankerEnabled && !reranker) {
-            // On but incomplete: search must keep working, so treat it as off.
-            log("Reranker enabled without a URL or model; searching without it", {
+        // On but incomplete, through the general switch or a search's own override:
+        // search must keep working, so treat the reranker as off and say so once.
+        if (!reranker &&
+            (options.reranker === true || (CONFIG.rerankerEnabled && options.reranker !== false))) {
+            log("Reranker asked for without a URL or model; searching without it", {
                 code: "reranker_not_configured",
             });
         }
@@ -99,6 +101,7 @@ export class LocalMemoryClient {
                     maxMemories,
                     similarityThreshold: CONFIG.similarityThreshold,
                     kind: options.kind ?? "normal",
+                    thresholdAfterRerank: options.thresholdAfterRerank,
                 }, reranker, {
                     search: (shardList, vector, tag, limit, threshold, queryText, kind) => tursoVectorSearch.searchAcrossShards(shardList, vector, tag, limit, threshold, queryText, keywordBoost, "boost", undefined, kind),
                     log: (message, meta) => log(message, meta),

package/dist/services/global-config-writer.js

@@ -56,6 +56,7 @@ const keys = new Set([
     "memorisationRepeat",
     "memorisationMaxSize",
     "memorisationSizeUnit",
+    "memorisationRerank",
 ]);
 function isFolder(path) {
     try {
@@ -117,6 +118,8 @@ function isValidEdit(key, value) {
         return value === null || KEYWORD_MODES.includes(value);
     if (key === "memorisationSizeUnit")
         return value === "characters" || value === "tokens";
+    if (key === "memorisationRerank")
+        return typeof value === "boolean";
     if (key === "memorisationRepeat" || key === "memorisationMaxSize")
         return true;
     // The opening context query: null removes it; an empty query means no opening search.

package/dist/services/linked-search.d.ts

@@ -19,6 +19,10 @@ export interface MemoryReadOptions {
     kind?: MemoryKind;
     /** Search only: the result limit for this read, instead of `maxMemories`. */
     limit?: number;
+    /** Search only: rerank this read whatever the general `rerankerEnabled` switch says. */
+    reranker?: boolean;
+    /** Search only: with the reranker on, still drop hits below `similarityThreshold`. */
+    thresholdAfterRerank?: boolean;
 }
 /** The settings a linked search needs; the caller snapshots them from CONFIG. */
 export interface LinkedSearchSettings {

package/dist/services/memorisation-routes.d.ts

@@ -1,16 +1,17 @@
-import type { MemoryScope } from "./client.js";
 import type { RouteResult } from "./chat/chat-routes.js";
 import type { LinkedSearchSettings } from "./linked-search.js";
 /** The CONFIG values recall needs. The route registration builds them, so this module has no CONFIG import. */
 export interface MemorisationRouteConfig {
-    defaultScope: MemoryScope | undefined;
     linked: LinkedSearchSettings;
+    /** The global `memorisationRerank` setting, so text recall may use the reranker. */
+    rerank: boolean;
 }
 /**
  * Routes for the Memorisation page. Returns undefined for paths it does not own.
  * `GET /api/memorisations?tags=a,b` lists, `GET /api/memorisations/recall?q=&tags=` recalls
- * and `POST /api/memorisations/bulk-delete` deletes. The general `/api/memories` delete
- * routes refuse memorisations by design, so this module owns the bulk delete. Memorisation
- * text is returned to the page and never written to the log.
+ * and `POST /api/memorisations/bulk-delete` deletes. Every read covers the whole active
+ * space and its linked spaces, so the page works from any folder. The general `/api/memories`
+ * delete routes refuse memorisations by design, so this module owns the bulk delete.
+ * Memorisation text is returned to the page and never written to the log.
  */
-export declare function routeMemorisations(path: string, method: string, url: URL, cwd: string, config: MemorisationRouteConfig, readBody?: () => Promise<any>): Promise<RouteResult | undefined>;
+export declare function routeMemorisations(path: string, method: string, url: URL, config: MemorisationRouteConfig, readBody?: () => Promise<any>): Promise<RouteResult | undefined>;

package/dist/services/memorisation-routes.js

@@ -11,28 +11,26 @@ const fail = (status, error) => ({
 /**
  * Routes for the Memorisation page. Returns undefined for paths it does not own.
  * `GET /api/memorisations?tags=a,b` lists, `GET /api/memorisations/recall?q=&tags=` recalls
- * and `POST /api/memorisations/bulk-delete` deletes. The general `/api/memories` delete
- * routes refuse memorisations by design, so this module owns the bulk delete. Memorisation
- * text is returned to the page and never written to the log.
+ * and `POST /api/memorisations/bulk-delete` deletes. Every read covers the whole active
+ * space and its linked spaces, so the page works from any folder. The general `/api/memories`
+ * delete routes refuse memorisations by design, so this module owns the bulk delete.
+ * Memorisation text is returned to the page and never written to the log.
  */
-export async function routeMemorisations(path, method, url, cwd, config, readBody) {
+export async function routeMemorisations(path, method, url, config, readBody) {
     const wanted = parseTagList(url.searchParams.get("tags") ?? undefined);
     if (method === "GET" && path === "/api/memorisations")
-        return listMemorisations(cwd, wanted);
+        return listMemorisations(wanted);
     if (method === "GET" && path === "/api/memorisations/recall") {
-        return recall(url.searchParams, wanted, cwd, config);
+        return recall(url.searchParams, wanted, config);
     }
     if (method === "POST" && path === "/api/memorisations/bulk-delete" && readBody) {
-        return bulkDelete(await readBody(), cwd, config);
+        return bulkDelete(await readBody(), config);
     }
     return undefined;
 }
-async function listMemorisations(cwd, wanted) {
-    const [{ listMemorisationsByTags }, { getTags }] = await Promise.all([
-        import("./memorisation-store.js"),
-        import("./tags.js"),
-    ]);
-    const result = await listMemorisationsByTags(getTags(cwd).project.tag, wanted, "project", LIST_LIMIT);
+async function listMemorisations(wanted) {
+    const { listMemorisationsByTags } = await import("./memorisation-store.js");
+    const result = await listMemorisationsByTags("", wanted, "all-projects", LIST_LIMIT);
     if (!result.success)
         return fail(500, result.error);
     const memorisations = result.memories.map((memory) => ({
@@ -46,20 +44,17 @@ async function listMemorisations(cwd, wanted) {
     }));
     return ok({ memorisations });
 }
-async function recall(params, wanted, cwd, config) {
+async function recall(params, wanted, config) {
     const text = params.get("q")?.trim();
     if (!text && wanted.length === 0)
         return fail(400, "q or tags required: an ID, text or labels");
     const limit = Number(params.get("limit"));
-    const [{ runRecall }, { getTags }] = await Promise.all([
-        import("../core/memorisation-operations.js"),
-        import("./tags.js"),
-    ]);
+    const { runRecall } = await import("../core/memorisation-operations.js");
     const result = await runRecall({
         query: text,
         tags: wanted.join(","),
         ...(Number.isInteger(limit) && limit > 0 ? { limit } : {}),
-    }, getTags(cwd), config.defaultScope, config.linked);
+    }, config.linked, config.rerank);
     if (result.success !== true) {
         // An unknown ID is a missing record. Any other refusal is a server problem.
         const error = String(result.error ?? "Recall failed");
@@ -67,8 +62,8 @@ async function recall(params, wanted, cwd, config) {
     }
     return ok({ count: result.count, matches: result.matches });
 }
-/** Deletes memorisations of this project from the active space. Linked spaces are never touched. */
-async function bulkDelete(body, cwd, config) {
+/** Deletes memorisations of any project from the active space. Linked spaces are never touched. */
+async function bulkDelete(body, config) {
     const ids = body?.ids;
     if (!Array.isArray(ids) ||
         ids.length < 1 ||
@@ -76,15 +71,11 @@ async function bulkDelete(body, cwd, config) {
         !ids.every((id) => typeof id === "string" && isMemorisationId(id))) {
         return fail(400, `ids must be 1 to ${BULK_DELETE_MAX} memorisation IDs (ms-xxxxxx)`);
     }
-    const [{ deleteMemorisation }, { getTags }] = await Promise.all([
-        import("./memorisation-store.js"),
-        import("./tags.js"),
-    ]);
-    const containerTag = getTags(cwd).project.tag;
+    const { deleteMemorisation } = await import("./memorisation-store.js");
     const deleted = [];
     const skipped = [];
     for (const id of new Set(ids)) {
-        const outcome = await deleteMemorisation(id, containerTag, config.linked);
+        const outcome = await deleteMemorisation(id, null, config.linked);
         if (outcome === "deleted")
             deleted.push(id);
         else

package/dist/services/memorisation-store.d.ts

@@ -10,11 +10,12 @@ export interface StoredMemorisation {
     space?: string;
 }
 /**
- * Looks a memorisation up by exact ID in the active space, then in the linked spaces. Only
- * rows of the given project are returned, the same rule as text recall and list.
+ * Looks a memorisation up by exact ID in the active space, then in the linked spaces. A
+ * project tag narrows the lookup to that project; `null` matches any project, the rule
+ * recall and bulk delete use.
  * The caller snapshots `linked` from CONFIG, so this module never imports it.
  */
-export declare function getMemorisationById(id: string, containerTag: string, linked: LinkedSearchSettings): Promise<StoredMemorisation | null>;
+export declare function getMemorisationById(id: string, containerTag: string | null, linked: LinkedSearchSettings): Promise<StoredMemorisation | null>;
 /**
  * True when any project in the active or a linked space already uses this memorisation ID.
  * A new ID must be free everywhere, so this check ignores the project.
@@ -38,8 +39,9 @@ export declare function listMemorisationsByTags(containerTag: string, tags: stri
 }>;
 export type MemorisationDeleteOutcome = "deleted" | "not-found" | "linked" | "not-memorisation";
 /**
- * Deletes one memorisation of this project from the active space. A row of another project,
- * or no row, is "not-found". A row of another type is kept ("not-memorisation"). A row that
- * lives only in a linked, read-only space is "linked" and is never touched.
+ * Deletes one memorisation from the active space. A project tag narrows the delete to that
+ * project; `null` matches any project, the rule bulk delete uses. No row is "not-found".
+ * A row of another type is kept ("not-memorisation"). A row that lives only in a linked,
+ * read-only space is "linked" and is never touched.
  */
-export declare function deleteMemorisation(id: string, containerTag: string, linked: LinkedSearchSettings): Promise<MemorisationDeleteOutcome>;
+export declare function deleteMemorisation(id: string, containerTag: string | null, linked: LinkedSearchSettings): Promise<MemorisationDeleteOutcome>;