om-memory-system 4.20.0 → 4.23.0-next.144

package/dist/adapters/opencode/profile-learning.d.ts

@@ -1,5 +1,4 @@
 import type { PluginInput } from "@opencode-ai/plugin";
-export declare function shouldRunAutomaticProfileCleanup(previousPromptCount: number, addedPromptCount: number, interval?: number): boolean;
 export declare function performUserProfileLearning(ctx: PluginInput, directory: string): Promise<void>;
 import { createUserProfileAnalysisSchema, USER_PROFILE_LLM_CONFIDENCE_MAX } from "../../core/extraction.js";
 export { createUserProfileAnalysisSchema, USER_PROFILE_LLM_CONFIDENCE_MAX };

package/dist/adapters/opencode/profile-learning.js

@@ -6,43 +6,11 @@ import { resolveOpencodeHostModel } from "../../services/ai/live-model-choice.js
 import { userPromptManager } from "../../services/user-prompt/user-prompt-manager.js";
 import { userProfileManager } from "../../services/user-profile/user-profile-manager.js";
 import { saveProfile } from "../../services/user-profile/merge-and-save.js";
+import { shouldStartAutoCleanup } from "../../services/user-profile/auto-cleanup-trigger.js";
 import { sortProfileItems } from "../../utils/profile.js";
 import { truncateToMaxBytes } from "../../utils/context-limit.js";
 import { loadOpencodeProvider } from "./opencode-provider-loader.js";
 let isLearningRunning = false;
-export function shouldRunAutomaticProfileCleanup(previousPromptCount, addedPromptCount, interval = CONFIG.userProfileAutoCleanupInterval) {
-    if (!CONFIG.userProfileAutoCleanupEnabled || !Number.isInteger(interval) || interval <= 0) {
-        return false;
-    }
-    return (Math.floor(previousPromptCount / interval) <
-        Math.floor((previousPromptCount + addedPromptCount) / interval));
-}
-async function runAutomaticProfileCleanup(userId) {
-    try {
-        const profile = await userProfileManager.getActiveProfile(userId);
-        if (!profile)
-            return;
-        const profileData = JSON.parse(profile.profileData);
-        const itemCount = profileData.preferences.length + profileData.patterns.length + profileData.workflows.length;
-        if (itemCount < 2)
-            return;
-        const { aiCleanupProfile } = await import("../../services/user-profile/ai-cleanup.js");
-        const result = await aiCleanupProfile(profileData);
-        if (result.diff.merged.length === 0 && result.diff.removed.length === 0)
-            return;
-        const summary = `Automatic AI cleanup: ${result.diff.merged.length} merged, ${result.diff.removed.length} removed`;
-        const updated = await saveProfile(profile, result.cleaned, summary);
-        log("user-profile-learning: automatic cleanup complete", {
-            userId,
-            updated,
-            merged: result.diff.merged.length,
-            removed: result.diff.removed.length,
-        });
-    }
-    catch (error) {
-        log("user-profile-learning: automatic cleanup failed", { userId, error: String(error) });
-    }
-}
 export async function performUserProfileLearning(ctx, directory) {
     if (isLearningRunning)
         return;
@@ -204,8 +172,14 @@ Rules:
             await userProfileManager.createProfile(userId, tags.user.displayName || "Unknown", tags.user.userName || "unknown", tags.user.userEmail || "unknown", llmResult, prompts.length);
             await userPromptManager.markMultipleAsUserLearningCaptured(prompts.map((p) => p.id));
         }
-        if (shouldRunAutomaticProfileCleanup(cleanupPreviousPromptCount, prompts.length)) {
-            await runAutomaticProfileCleanup(userId);
+        if (shouldStartAutoCleanup(cleanupPreviousPromptCount, prompts.length, {
+            enabled: CONFIG.userProfileAutoCleanupEnabled,
+            interval: CONFIG.userProfileAutoCleanupInterval,
+        })) {
+            // In the background: a long run must not hold the learning pass, and a failure must not fail it.
+            void import("../../services/user-profile/auto-cleanup.js")
+                .then(({ startAutoProfileCleanup }) => startAutoProfileCleanup(userId))
+                .catch(() => { });
         }
         if (CONFIG.showUserProfileToasts) {
             await ctx.client?.tui

package/dist/adapters/pi/profile.js

@@ -3,6 +3,7 @@ import { analyzeProfile, } from "../../core/profile-analysis.js";
 import { profileBackoff, recordProfileFailure } from "../../core/profile-backoff.js";
 import { log } from "../../services/logger.js";
 import { getTags } from "../../services/tags.js";
+import { shouldStartAutoCleanup } from "../../services/user-profile/auto-cleanup-trigger.js";
 function replyText(reply) {
     return reply.content
         .filter((block) => block.type === "text" && typeof block.text === "string")
@@ -37,6 +38,13 @@ export function createPiProfileAnalyzer(resolveModel, timeoutMs = 120000) {
         },
     };
 }
+/** The model Pi uses for profile work, for the automatic cleanup. */
+function resolveCleanupModel(input) {
+    if (input.resolveProfileModel)
+        return input.resolveProfileModel();
+    const handle = input.resolveModel?.();
+    return handle ? adaptPiProfileModel(handle) : null;
+}
 /**
  * Pi-side profile learning lifecycle: analyse a batch of settled user prompts
  * through the model bridge and persist via the shared profile manager. Failures
@@ -95,6 +103,15 @@ export async function performPiProfileLearning(input) {
             prompts: input.prompts.length,
             hadExisting: Boolean(existing),
         });
+        if (shouldStartAutoCleanup(existing?.totalPromptsAnalyzed ?? 0, input.prompts.length, {
+            enabled: CONFIG.userProfileAutoCleanupEnabled,
+            interval: CONFIG.userProfileAutoCleanupInterval,
+        })) {
+            // In the background: a long run must not hold the learning pass, and a failure must not fail it.
+            void import("../../services/user-profile/auto-cleanup.js")
+                .then(({ startAutoProfileCleanup }) => startAutoProfileCleanup(userId, { resolveModel: () => resolveCleanupModel(input) }))
+                .catch(() => { });
+        }
         if (CONFIG.showUserProfileToasts && input.notify) {
             await Promise.resolve(input.notify({
                 title: "User Profile Updated",

package/dist/config.d.ts

@@ -107,7 +107,7 @@ interface OmmsConfig {
         injectOn?: "first" | "always";
     };
 }
-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  // 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  // Periodically merge duplicate or irrelevant profile items with the configured AI provider\n  \"userProfileAutoCleanupEnabled\": true,\n  // Number of analyzed user prompts between automatic AI cleanup runs\n  \"userProfileAutoCleanupInterval\": 100,\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\n  \"similarityThreshold\": 0.6,\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  // Recent memories added at fresh session start (OpenCode V1 and Claude Code).\n  // Pi and OpenCode V2 search the store with each prompt instead.\n  \"chatMessage\": {\n    \"enabled\": true,\n    // Recent memories injected at session start.\n    \"maxMemories\": 3,\n    \"excludeCurrentSession\": true,\n    // When to inject: \"first\" message of a session or \"always\".\n    // \"injectOn\": \"first\",\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  // ============================================\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  // 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\n  \"similarityThreshold\": 0.6,\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  // Recent memories added at fresh session start (OpenCode V1 and Claude Code).\n  // Pi and OpenCode V2 search the store with each prompt instead.\n  \"chatMessage\": {\n    \"enabled\": true,\n    // Recent memories injected at session start.\n    \"maxMemories\": 3,\n    \"excludeCurrentSession\": true,\n    // When to inject: \"first\" message of a session or \"always\".\n    // \"injectOn\": \"first\",\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  // ============================================\n  // Advanced Settings\n  // ============================================\n  \n  // Inject user profile into AI context (preferences, patterns, workflows)\n  \"injectProfile\": true\n}\n";
 export declare function normalizeAutoCaptureMaxContextBytes(value: number): number;
 /** Whole hours from 0 to 720. A value that is not a number gives the default. */
 export declare function normalizeCaptureRetryRetentionHours(value: unknown): number;

package/dist/config.js

@@ -92,7 +92,7 @@ const DEFAULTS = {
     userProfileEmbeddingMinDescriptionLength: 5,
     userProfileMinEvidenceForRetention: 3,
     userProfileAutoCleanupEnabled: true,
-    userProfileAutoCleanupInterval: 100,
+    userProfileAutoCleanupInterval: 500,
     userProfileValidationEnabled: false,
     showAutoCaptureToasts: true,
     showUserProfileToasts: true,
@@ -486,10 +486,9 @@ export const CONFIG_TEMPLATE = `{
   // Items confirmed fewer times are more likely to be pruned when confidence decays
   "userProfileMinEvidenceForRetention": 3,
 
-  // Periodically merge duplicate or irrelevant profile items with the configured AI provider
+  // Merge duplicate or vague profile items every N analysed prompts (about 1 model call per 100 items)
   "userProfileAutoCleanupEnabled": true,
-  // Number of analyzed user prompts between automatic AI cleanup runs
-  "userProfileAutoCleanupInterval": 100,
+  "userProfileAutoCleanupInterval": 500,
 
   // Enable LLM validation of existing preferences against recent behavior.
   // When enabled, each analysis round checks if top-5 preferences still match recent prompts.
@@ -872,6 +871,8 @@ export function initConfig(directory, options = {}) {
     const projectOverrides = { ...projectConfig };
     delete projectOverrides.autoCleanupEnabled;
     delete projectOverrides.autoCleanupRetentionDays;
+    delete projectOverrides.userProfileAutoCleanupEnabled;
+    delete projectOverrides.userProfileAutoCleanupInterval;
     // A checked-in project file must not start recording conversations; it may
     // only turn tracing off for its project.
     if (projectOverrides.captureTrace === true) {

package/dist/importer/claude-hook-api.js

@@ -14,6 +14,7 @@ import { log } from "../services/logger.js";
 import { isFullyPrivate, stripPrivateContent } from "../services/privacy.js";
 import { getTags } from "../services/tags.js";
 import { extractClaudeWindowsAfter, parseClaudeTranscript, } from "./claude-conversation.js";
+import { learnClaudeProfile } from "./claude-profile-learning.js";
 import { defaultClaudeProjectsRoot } from "./claude-reader.js";
 // The web app's side of the Claude Code hooks (design decisions 4, 6, 7):
 // retrieval for SessionStart and UserPromptSubmit, a one-worker capture queue
@@ -388,8 +389,9 @@ async function captureTurn(request, deps) {
         // The model answers again, so earlier failed turns may go through now.
         startCaptureRetryDrain(HOST);
     }
-    if (counts.captured + counts.skipped > 0)
-        await learnProfile(request.cwd, deps);
+    if (counts.captured + counts.skipped > 0) {
+        await learnClaudeProfile(request.cwd, async () => deps.profileModel?.() ?? (await externalModels()).profile);
+    }
 }
 // --- Profile learning ---
 async function recordPrompt(request, unit, memoryId) {
@@ -407,70 +409,6 @@ async function recordPrompt(request, unit, memoryId) {
         });
     }
 }
-let learning = false;
-/**
- * Build or update the user profile from the recorded prompts once there are
- * `userProfileAnalysisInterval` of them. A failure is logged and never
- * blocks capture; the prompts stay for the next pass.
- */
-async function learnProfile(directory, deps) {
-    const interval = CONFIG.userProfileAnalysisInterval ?? 0;
-    if (!(interval > 0) || learning)
-        return;
-    const { profileBackoff, profileCatchUpLock, recordProfileFailure } = await import("../core/profile-backoff.js");
-    // A catch-up run, here or in a terminal, analyses the same prompts; skip the live pass meanwhile.
-    if (!profileBackoff.canRun() || profileCatchUpLock.isActive())
-        return;
-    const { CatchUpLease } = await import("../services/user-prompt/profile-catch-up-lease.js");
-    if (await new CatchUpLease().isActive().catch(() => false))
-        return;
-    learning = true;
-    try {
-        const [{ userPromptManager }, { userProfileManager }, { analyzeProfile }] = await Promise.all([
-            import("../services/user-prompt/user-prompt-manager.js"),
-            import("../services/user-profile/user-profile-manager.js"),
-            import("../core/profile-analysis.js"),
-        ]);
-        await userPromptManager.skipTrivialPromptsForLearning();
-        if ((await userPromptManager.countUnanalyzedForUserLearning()) < interval)
-            return;
-        const user = getTags(directory).user;
-        if (!user.userEmail) {
-            log("Claude Code profile learning skipped", { code: "no-user-email" });
-            return;
-        }
-        const batch = await userPromptManager.getPromptsForUserLearning(interval, {
-            recentFirst: true,
-        });
-        if (batch.length === 0)
-            return;
-        const model = deps.profileModel?.() ?? (await externalModels()).profile;
-        const existing = await userProfileManager.getActiveProfile(user.userEmail);
-        const context = batch.map((prompt, index) => `${index + 1}. ${prompt.content}`).join("\n");
-        const analysis = await analyzeProfile(model, context, existing ? { id: existing.id, profileData: existing.profileData } : null);
-        if (existing) {
-            const { saveMergedProfile } = await import("../services/user-profile/merge-and-save.js");
-            const updated = await saveMergedProfile(userProfileManager, existing, analysis, batch.length, `Claude Code analysis of ${batch.length} prompts`);
-            if (!updated)
-                throw new Error("Profile update conflict");
-        }
-        else {
-            await userProfileManager.createProfile(user.userEmail, user.displayName || user.userEmail, user.userName || user.userEmail, user.userEmail, analysis.raw, batch.length);
-        }
-        await userPromptManager.markMultipleAsUserLearningCaptured(batch.map((prompt) => prompt.id));
-        profileBackoff.recordSuccess();
-        log("Claude Code profile learning finished", {
-            prompts: batch.length,
-            hadExisting: Boolean(existing),
-        });
-    }
-    catch (error) {
-        recordProfileFailure("claude-code", "Claude Code profile learning failed", error);
-    }
-    finally {
-        learning = false;
-    }
-}
 // --- Web app start ---
 /**
  * Register the Claude Code retry drain in this web app and start one retry

package/dist/importer/claude-profile-learning.d.ts

@@ -0,0 +1,7 @@
+import type { ModelPort } from "../core/profile-analysis.js";
+/**
+ * Build or update the user profile from the recorded prompts once there are
+ * `userProfileAnalysisInterval` of them. A failure is logged and never
+ * blocks capture; the prompts stay for the next pass.
+ */
+export declare function learnClaudeProfile(directory: string, resolveModel: () => ModelPort | Promise<ModelPort>): Promise<void>;

package/dist/importer/claude-profile-learning.js

@@ -0,0 +1,78 @@
+import { CONFIG } from "../config.js";
+import { log } from "../services/logger.js";
+import { getTags } from "../services/tags.js";
+import { shouldStartAutoCleanup } from "../services/user-profile/auto-cleanup-trigger.js";
+let learning = false;
+/**
+ * Build or update the user profile from the recorded prompts once there are
+ * `userProfileAnalysisInterval` of them. A failure is logged and never
+ * blocks capture; the prompts stay for the next pass.
+ */
+export async function learnClaudeProfile(directory, resolveModel) {
+    const interval = CONFIG.userProfileAnalysisInterval ?? 0;
+    if (!(interval > 0) || learning)
+        return;
+    const { profileBackoff, profileCatchUpLock, recordProfileFailure } = await import("../core/profile-backoff.js");
+    // A catch-up run, here or in a terminal, analyses the same prompts; skip the live pass meanwhile.
+    if (!profileBackoff.canRun() || profileCatchUpLock.isActive())
+        return;
+    const { CatchUpLease } = await import("../services/user-prompt/profile-catch-up-lease.js");
+    if (await new CatchUpLease().isActive().catch(() => false))
+        return;
+    learning = true;
+    try {
+        const [{ userPromptManager }, { userProfileManager }, { analyzeProfile }] = await Promise.all([
+            import("../services/user-prompt/user-prompt-manager.js"),
+            import("../services/user-profile/user-profile-manager.js"),
+            import("../core/profile-analysis.js"),
+        ]);
+        await userPromptManager.skipTrivialPromptsForLearning();
+        if ((await userPromptManager.countUnanalyzedForUserLearning()) < interval)
+            return;
+        const user = getTags(directory).user;
+        if (!user.userEmail) {
+            log("Claude Code profile learning skipped", { code: "no-user-email" });
+            return;
+        }
+        const batch = await userPromptManager.getPromptsForUserLearning(interval, {
+            recentFirst: true,
+        });
+        if (batch.length === 0)
+            return;
+        const model = await resolveModel();
+        const existing = await userProfileManager.getActiveProfile(user.userEmail);
+        const context = batch.map((prompt, index) => `${index + 1}. ${prompt.content}`).join("\n");
+        const analysis = await analyzeProfile(model, context, existing ? { id: existing.id, profileData: existing.profileData } : null);
+        if (existing) {
+            const { saveMergedProfile } = await import("../services/user-profile/merge-and-save.js");
+            const updated = await saveMergedProfile(userProfileManager, existing, analysis, batch.length, `Claude Code analysis of ${batch.length} prompts`);
+            if (!updated)
+                throw new Error("Profile update conflict");
+        }
+        else {
+            await userProfileManager.createProfile(user.userEmail, user.displayName || user.userEmail, user.userName || user.userEmail, user.userEmail, analysis.raw, batch.length);
+        }
+        await userPromptManager.markMultipleAsUserLearningCaptured(batch.map((prompt) => prompt.id));
+        profileBackoff.recordSuccess();
+        log("Claude Code profile learning finished", {
+            prompts: batch.length,
+            hadExisting: Boolean(existing),
+        });
+        if (shouldStartAutoCleanup(existing?.totalPromptsAnalyzed ?? 0, batch.length, {
+            enabled: CONFIG.userProfileAutoCleanupEnabled,
+            interval: CONFIG.userProfileAutoCleanupInterval,
+        })) {
+            // In the background, with the model learning used: a long run must not hold the pass.
+            const userId = user.userEmail;
+            void import("../services/user-profile/auto-cleanup.js")
+                .then(({ startAutoProfileCleanup }) => startAutoProfileCleanup(userId, { resolveModel: () => model }))
+                .catch(() => { });
+        }
+    }
+    catch (error) {
+        recordProfileFailure("claude-code", "Claude Code profile learning failed", error);
+    }
+    finally {
+        learning = false;
+    }
+}

package/dist/services/ai-cleanup-handlers.js

@@ -95,7 +95,7 @@ export async function handleAiCleanupApply(userId, body, deps = {}) {
     try {
         const { getAiCleanupResult, stopAiCleanupJob } = await import("./user-profile/ai-cleanup-job.js");
         const { applyCleanupEdits } = await import("./user-profile/ai-cleanup-apply.js");
-        const { MAX_PROFILE_SAVE_ATTEMPTS, saveProfile } = await import("./user-profile/merge-and-save.js");
+        const { applyToNewestProfile } = await import("./user-profile/profile-edit-save.js");
         const user = await cleanupUser(userId);
         const job = getAiCleanupResult(user);
         if (!job)
@@ -104,21 +104,21 @@ export async function handleAiCleanupApply(userId, body, deps = {}) {
             merged: Array.isArray(body?.acceptedMerged) ? body.acceptedMerged : [],
             removed: Array.isArray(body?.acceptedRemoved) ? body.acceptedRemoved : [],
         };
-        const manager = await managerOf(deps);
-        for (let attempt = 1; attempt <= MAX_PROFILE_SAVE_ATTEMPTS; attempt++) {
-            const profile = await manager.getActiveProfile(user);
-            if (!profile)
-                return { success: false, error: "Profile not found" };
-            const result = applyCleanupEdits(JSON.parse(profile.profileData), job, accepted);
-            const saved = result.applied === 0 ||
-                (await saveProfile(profile, result.profile, "AI cleanup applied", 0, manager));
-            if (saved) {
-                stopAiCleanupJob(user);
-                return { success: true, data: { applied: result.applied, skipped: result.skipped } };
-            }
-            log("AI cleanup apply: save conflict", { attempt, of: MAX_PROFILE_SAVE_ATTEMPTS });
+        const outcome = await applyToNewestProfile(user, (data) => {
+            const applied = applyCleanupEdits(data, job, accepted);
+            return {
+                profile: applied.profile,
+                changed: applied.applied > 0,
+                result: { applied: applied.applied, skipped: applied.skipped },
+            };
+        }, { summary: "AI cleanup applied", store: await managerOf(deps) });
+        if (outcome.status === "no-profile")
+            return { success: false, error: "Profile not found" };
+        if (outcome.status === "conflict") {
+            return { success: false, error: "Profile saved at the same moment. Try Apply again." };
         }
-        return { success: false, error: "Profile saved at the same moment. Try Apply again." };
+        stopAiCleanupJob(user);
+        return { success: true, data: outcome.result };
     }
     catch (error) {
         return handlerError("apply", error);

package/dist/services/auto-cleanup-handlers.d.ts

@@ -0,0 +1,22 @@
+import type { UserProfileManager } from "./user-profile/user-profile-manager.js";
+import type { AutoCleanupSummary } from "./user-profile/auto-cleanup-store.js";
+interface ApiResponse<T = any> {
+    success: boolean;
+    data?: T;
+    error?: string;
+}
+export interface AutoCleanupHandlerDeps {
+    manager?: Pick<UserProfileManager, "getActiveProfile" | "updateProfile">;
+}
+/** The newest automatic cleanup that can still be undone, without its items. */
+export declare function handleAutoCleanupStatus(userId?: string, deps?: AutoCleanupHandlerDeps): Promise<ApiResponse<AutoCleanupSummary | null>>;
+/**
+ * Undoes the newest automatic cleanup that is not undone, in the newest profile, with the
+ * version check. A save conflict is tried again at most three times; after that the profile
+ * and the record stay as they were.
+ */
+export declare function handleAutoCleanupUndo(userId?: string, deps?: AutoCleanupHandlerDeps): Promise<ApiResponse<{
+    restored: number;
+    notFound: number;
+}>>;
+export {};

package/dist/services/auto-cleanup-handlers.js

@@ -0,0 +1,71 @@
+import { log } from "./logger.js";
+async function cleanupUser(userId) {
+    if (userId)
+        return userId;
+    const { CONFIG } = await import("../config.js");
+    const { resolveWebProfileUserId } = await import("./profile-identity.js");
+    return (await resolveWebProfileUserId(CONFIG)) || "unknown";
+}
+async function managerOf(deps) {
+    return (deps.manager ?? (await import("./user-profile/user-profile-manager.js")).userProfileManager);
+}
+function handlerError(step, error) {
+    // The log never holds profile text, so only the error's kind is kept.
+    log(`Automatic cleanup ${step}: error`, {
+        code: error instanceof Error ? error.name : "unknown",
+    });
+    return { success: false, error: String(error) };
+}
+/** The newest automatic cleanup that can still be undone, without its items. */
+export async function handleAutoCleanupStatus(userId, deps = {}) {
+    try {
+        const { latestOpenAutoCleanupSummary } = await import("./user-profile/auto-cleanup-store.js");
+        const profile = await (await managerOf(deps)).getActiveProfile(await cleanupUser(userId));
+        if (!profile)
+            return { success: true, data: null };
+        return { success: true, data: await latestOpenAutoCleanupSummary(profile.id) };
+    }
+    catch (error) {
+        return handlerError("status", error);
+    }
+}
+/**
+ * Undoes the newest automatic cleanup that is not undone, in the newest profile, with the
+ * version check. A save conflict is tried again at most three times; after that the profile
+ * and the record stay as they were.
+ */
+export async function handleAutoCleanupUndo(userId, deps = {}) {
+    try {
+        const { latestOpenAutoCleanup, markAutoCleanupUndone } = await import("./user-profile/auto-cleanup-store.js");
+        const { undoCleanupEdits } = await import("./user-profile/ai-cleanup-undo.js");
+        const { applyToNewestProfile } = await import("./user-profile/profile-edit-save.js");
+        const user = await cleanupUser(userId);
+        const manager = await managerOf(deps);
+        const profile = await manager.getActiveProfile(user);
+        const record = profile ? await latestOpenAutoCleanup(profile.id) : null;
+        if (!record)
+            return { success: false, error: "Nothing to undo." };
+        const outcome = await applyToNewestProfile(user, (data) => {
+            const undone = undoCleanupEdits(data, record.entries);
+            return {
+                profile: undone.profile,
+                changed: undone.restored > 0,
+                result: { restored: undone.restored, notFound: undone.notFound },
+            };
+        }, { summary: "Automatic cleanup undone", store: manager });
+        if (outcome.status === "no-profile")
+            return { success: false, error: "Nothing to undo." };
+        if (outcome.status === "conflict") {
+            return { success: false, error: "Profile saved at the same moment. Try Undo again." };
+        }
+        await markAutoCleanupUndone(record.id);
+        log("Automatic cleanup undone", {
+            restored: outcome.result.restored,
+            notFound: outcome.result.notFound,
+        });
+        return { success: true, data: outcome.result };
+    }
+    catch (error) {
+        return handlerError("undo", error);
+    }
+}

package/dist/services/cleanup-service.js

@@ -95,6 +95,15 @@ export class CleanupService {
                     code: error instanceof Error ? error.name : "unknown",
                 });
             }
+            try {
+                const { pruneAutoCleanupRecords } = await import("./user-profile/auto-cleanup-store.js");
+                await pruneAutoCleanupRecords();
+            }
+            catch (error) {
+                log("Automatic cleanup record cleanup failed", {
+                    code: error instanceof Error ? error.name : "unknown",
+                });
+            }
             try {
                 await userPromptManager.vacuum();
                 log("Cleanup: VACUUM done", { db: "user-prompts.db" });

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

@@ -18,6 +18,8 @@ const keys = new Set([
     "opencodeBackfillModel",
     "piBackfillModel",
     "webServerAutoStart",
+    "userProfileAutoCleanupEnabled",
+    "userProfileAutoCleanupInterval",
     "memoryProvider",
     "memoryApiUrl",
     "memoryModel",
@@ -25,6 +27,7 @@ const keys = new Set([
     "importPathMaps",
     "importIgnoredDirectories",
     "claudeConfigDir",
+    "autoCaptureLanguage",
     "maxMemories",
     "chatMessage.maxMemories",
     "autoCaptureMaxContextBytes",
@@ -69,9 +72,17 @@ function isValidEdit(key, value) {
     if (key === "embeddingDimensions") {
         return Number.isInteger(value) && value > 0 && value <= 65536;
     }
-    if (key === "captureTrace" || key === "autoBackfill" || key === "webServerAutoStart") {
+    if (key === "captureTrace" ||
+        key === "autoBackfill" ||
+        key === "webServerAutoStart" ||
+        key === "userProfileAutoCleanupEnabled") {
         return typeof value === "boolean";
     }
+    // Each automatic cleanup is paid for, so the page refuses an interval that would run one
+    // after almost every learning pass. A hand edit of the file has no floor.
+    if (key === "userProfileAutoCleanupInterval") {
+        return Number.isSafeInteger(value) && value >= 50 && value <= 100000;
+    }
     if (key.endsWith("RetentionDays"))
         return Number.isSafeInteger(value) && value >= 1;
     if (key === "captureRetryRetentionHours") {
@@ -81,6 +92,9 @@ function isValidEdit(key, value) {
         return isSecretReference(value);
     if (key === "memoryProvider")
         return MEMORY_PROVIDERS.includes(value);
+    // "auto", or an ISO 639 code (en, id, zh, fil).
+    if (key === "autoCaptureLanguage")
+        return typeof value === "string" && /^(auto|[a-z]{2,3})$/.test(value);
     // Entries are checked by the startup validation below.
     if (key === "importPathMaps" || key === "importIgnoredDirectories")
         return Array.isArray(value);