openclaw-hybrid-memory 2026.3.293 → 2026.3.301

package/cli/cmd-verify.ts

@@ -368,11 +368,14 @@ export async function runVerifyForCli(
           model: modelForTest,
           dimensions: dimensionsForTest,
           batchSize: cfg.embedding.batchSize ?? 32,
+          ...(typeof cfg.embedding.deployment === "string" && cfg.embedding.deployment.trim()
+            ? { deployment: cfg.embedding.deployment.trim() }
+            : {}),
+          ...(cfg.embedding.models?.length ? { models: cfg.embedding.models } : {}),
           ...(p === "openai" && {
             apiKey: cfg.embedding.apiKey,
-            ...(typeof (cfg.embedding as Record<string, unknown>).endpoint === "string" &&
-            (cfg.embedding as Record<string, unknown>).endpoint
-              ? { endpoint: (cfg.embedding as Record<string, unknown>).endpoint as string }
+            ...(typeof cfg.embedding.endpoint === "string" && cfg.embedding.endpoint.trim()
+              ? { endpoint: cfg.embedding.endpoint.trim() }
               : {}),
           }),
           ...(p === "google" && {

package/lifecycle/hooks.ts

@@ -22,6 +22,7 @@ import { registerFrustrationHandlers } from "./stage-frustration.js";
 import { createSessionState } from "./session-state.js";
 import type { LifecycleContext, SessionState } from "./types.js";
 import { capturePluginError } from "../services/error-reporter.js";
+import { isAbortOrTransientLlmError } from "../services/chat.js";
 import { buildDailyNarrative } from "../src/worker/narratives.js";
 
 export type { LifecycleContext } from "./types.js";
@@ -155,12 +156,20 @@ export function createLifecycleHooks(ctx: LifecycleContext) {
           fallbackModels: [],
         });
       } catch (err) {
-        capturePluginError(err instanceof Error ? err : new Error(String(err)), {
-          subsystem: "narratives",
-          operation: "agent-end-build-narrative",
-          sessionId,
-        });
-        api.logger.warn(`memory-hybrid: session narrative build failed: ${String(err)}`);
+        const transient = isAbortOrTransientLlmError(err);
+        if (!transient) {
+          capturePluginError(err instanceof Error ? err : new Error(String(err)), {
+            subsystem: "narratives",
+            operation: "agent-end-build-narrative",
+            sessionId,
+          });
+        }
+        const detail = err instanceof Error ? err.message : String(err);
+        if (transient) {
+          api.logger.info?.(`memory-hybrid: session narrative skipped (LLM unavailable or aborted): ${detail}`);
+        } else {
+          api.logger.warn(`memory-hybrid: session narrative build failed: ${String(err)}`);
+        }
       }
     });
   };

package/lifecycle/stage-recall.ts

@@ -18,6 +18,7 @@ import {
 } from "../services/ambient-retrieval.js";
 import { capturePluginError } from "../services/error-reporter.js";
 import { formatNarrativeRange, recallNarrativeSummaries } from "../services/narrative-recall.js";
+import { yieldEventLoop } from "../utils/event-loop-yield.js";
 import { withTimeout } from "../utils/timeout.js";
 import { estimateTokens } from "../utils/text.js";
 import { isConsolidatedDerivedFact } from "../utils/consolidation-controls.js";
@@ -64,6 +65,9 @@ async function runRecall(
 
     api.logger.debug?.(`memory-hybrid: auto-recall start (prompt length ${e.prompt.length})`);
 
+    // Let pending gateway I/O (health RPCs, WebSocket) run before heavy sync work (#931).
+    await yieldEventLoop();
+
     const fmt = ctx.cfg.autoRecall.injectionFormat;
     const isProgressive = fmt === "progressive" || fmt === "progressive_hybrid";
     const searchLimit = isProgressive
@@ -110,6 +114,7 @@ async function runRecall(
       };
       const degradedLimit = ctx.cfg.autoRecall.limit;
       const trimmed = e.prompt.trim();
+      await yieldEventLoop();
       const ftsOnly = ctx.factsDb.search(trimmed, degradedLimit, recallOpts);
       let hotPart = "";
       if (ctx.cfg.memoryTiering.enabled && ctx.cfg.memoryTiering.hotMaxTokens > 0) {
@@ -218,6 +223,7 @@ async function runRecall(
         procedureBlock = block;
       }
     }
+    await yieldEventLoop();
     const withProcedures = (s: string) => (procedureBlock ? `${procedureBlock}\n${s}` : s);
 
     // HOT block
@@ -233,6 +239,8 @@ async function runRecall(
       }
     }
 
+    await yieldEventLoop();
+
     const recallOpts = {
       tierFilter,
       scopeFilter,
@@ -315,6 +323,7 @@ async function runRecall(
         if (extraQueries.length > 0) {
           const extraResultSets: SearchResult[][] = [candidates];
           for (const q of extraQueries) {
+            await yieldEventLoop();
             try {
               const qResults = await runRecallPipelineQuery(q.text, Math.ceil(limit / 2), pipelineDeps, hydeUsedRef, {
                 entity: q.type === "entity" ? q.entity : undefined,
@@ -398,6 +407,8 @@ async function runRecall(
       }
     }
 
+    await yieldEventLoop();
+
     const promptLower = e.prompt.toLowerCase();
     const { entityLookup } = ctx.cfg.autoRecall;
     if (entityLookup.enabled && entityLookup.entities.length > 0) {

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "openclaw-hybrid-memory",
-  "version": "2026.3.293",
+  "version": "2026.3.301",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "openclaw-hybrid-memory",
-      "version": "2026.3.293",
+      "version": "2026.3.301",
       "hasInstallScript": true,
       "dependencies": {
         "@lancedb/lancedb": "^0.27.1",

package/openclaw.plugin.json

@@ -1,7 +1,7 @@
 {
   "id": "openclaw-hybrid-memory",
   "kind": "memory",
-  "version": "2026.3.293",
+  "version": "2026.3.301",
   "uiHints": {
     "embedding.provider": {
       "label": "Embedding Provider",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openclaw-hybrid-memory",
-  "version": "2026.3.293",
+  "version": "2026.3.301",
   "type": "module",
   "description": "Give your OpenClaw agent lasting memory: structured facts, semantic search, auto-capture & recall, decay, optional credential vault. Part of Hybrid Memory v3.",
   "files": [

package/services/chat.ts

@@ -413,6 +413,7 @@ export async function chatComplete(opts: {
     const msg = error.message.toLowerCase();
     const isTransient =
       msg.includes("request was aborted") ||
+      msg.includes("operation was aborted") ||
       msg.includes("request timed out") ||
       msg.includes("timed out") ||
       msg.includes("llm request timeout") || // #339: our own timeout message uses "timeout" not "timed out"
@@ -464,6 +465,28 @@ export class LLMRetryError extends Error {
   }
 }
 
+/**
+ * True when an LLM call failed for abort, gateway loss, or transport — not plugin logic.
+ * Used by session narrative and similar paths to avoid noisy warns when the gateway stops.
+ */
+export function isAbortOrTransientLlmError(err: unknown): boolean {
+  if (err instanceof LLMRetryError) {
+    return isAbortOrTransientLlmError(err.cause);
+  }
+  if (err && typeof err === "object" && "cause" in err) {
+    const c = (err as { cause?: unknown }).cause;
+    if (c !== undefined && c !== null && isAbortOrTransientLlmError(c)) return true;
+  }
+  if (!(err instanceof Error)) {
+    return isConnectionErrorLike(err);
+  }
+  if (err.name === "AbortError") return true;
+  const msg = err.message;
+  if (/request was aborted|Request was aborted|The operation was aborted|operation was aborted/i.test(msg)) return true;
+  if (/gateway client stopped|gateway not reachable|not reachable\.|is it running/i.test(msg)) return true;
+  return isConnectionErrorLike(err);
+}
+
 /**
  * Retry wrapper for LLM calls with exponential backoff.
  * Retries on failure with increasing delays: 1s, 3s, 9s.
@@ -726,10 +749,8 @@ export async function chatCompleteWithRetry(opts: {
   const finalIsOOM = isOllamaOOM(finalError); // #387: OOM is expected when model too large for RAM
   const finalIs429 = is429OrWrapped(finalError); // #397
   const finalIsContextLength = isContextLengthError(finalError); // #488: input too long for model context window
-  const finalIsTimeout = /timed out|llm request timeout|request was aborted|Request was aborted/i.test(
-    finalError.message,
-  );
-  const finalIsConnectionError = isConnectionErrorLike(finalError);
+  /** Unwraps LLMRetryError so "Request was aborted" in the cause is detected (#935, #936). */
+  const finalIsTransientLlm = isAbortOrTransientLlmError(finalError);
 
   // When every model failed because provider keys are missing, queue a user-visible chat warning
   // and skip Sentry (this is a config issue, not a bug).
@@ -755,8 +776,7 @@ export async function chatCompleteWithRetry(opts: {
       !finalIsOOM &&
       !finalIsContextLength && // #488: context window exceeded = config issue, not a bug
       !finalIsUnconfigured &&
-      !finalIsTimeout &&
-      !finalIsConnectionError &&
+      !finalIsTransientLlm &&
       !finalIs403 &&
       !finalIs401 &&
       !finalIs429
@@ -803,10 +823,8 @@ export async function chatCompleteWithRetry(opts: {
       "⚠️ Memory plugin: LLM unauthorized (401) — your API key is invalid or expired. Check provider settings. " +
         "Run: openclaw hybrid-mem verify --test-llm",
     );
-  } else if (finalIsTimeout) {
-    // #339: timeout errors are transient — don't report to GlitchTip
-  } else if (finalIsConnectionError) {
-    // #703: OpenAI SDK "Connection error." / APIConnectionError is transient — don't report to GlitchTip
+  } else if (finalIsTransientLlm) {
+    // #339, #703, #935, #936: abort/timeout/connection (including LLMRetryError-wrapped causes) — don't report
   } else if (finalIs429) {
     // #397: rate limit / usage limit — transient provider error, don't report to GlitchTip
     pendingWarnings?.add(

package/services/embeddings/factory.ts

@@ -11,6 +11,7 @@ import {
   GOOGLE_EMBED_DEFAULT_MODEL,
   KNOWN_GOOGLE_EMBED_MODELS,
   OPENAI_ONLY_EMBED_MODELS,
+  isAzureOpenAiCompatibleEndpoint,
   isAzureOpenAiResourceEndpoint,
 } from "./shared.js";
 import { Embeddings } from "./openai-provider.js";
@@ -84,6 +85,24 @@ function openaiEmbeddingClientOpts(
   return opts;
 }
 
+/** Local/Ollama/ONNX embedding ids — never use as OpenAI/Azure `model` when falling back from Ollama/ONNX (#932). */
+function isLocalOnlyEmbeddingModelId(model: string | undefined): boolean {
+  if (!model) return false;
+  switch (model) {
+    case "nomic-embed-text":
+    case "mxbai-embed-large":
+    case "bge-m3":
+    case "bge-large":
+    case "bge-small-en-v1.5":
+    case "snowflake-arctic-embed":
+    case "all-minilm":
+    case "all-MiniLM-L6-v2":
+      return true;
+    default:
+      return false;
+  }
+}
+
 /** API model id(s) for OpenAI-compatible embeddings: optional Azure deployment name overrides logical `model`. */
 function openAiEmbeddingApiModels(cfg: EmbeddingConfig, forFallback = false): string[] {
   const { model, models, deployment } = cfg;
@@ -94,6 +113,12 @@ function openAiEmbeddingApiModels(cfg: EmbeddingConfig, forFallback = false): st
     if (model && OPENAI_ONLY_EMBED_MODELS.has(model)) {
       return [model];
     }
+    // Fallback path used for chain OpenAI arm and Ollama/ONNX→OpenAI fallback. Azure deployment names
+    // are often not in OPENAI_ONLY_EMBED_MODELS; do not substitute text-embedding-3-small (#932).
+    const m = typeof model === "string" ? model.trim() : "";
+    if (m && isAzureOpenAiCompatibleEndpoint(cfg.endpoint) && !isLocalOnlyEmbeddingModelId(m)) {
+      return [m];
+    }
     return ["text-embedding-3-small"];
   }
   return models?.length ? models : [model];

package/services/embeddings/shared.ts

@@ -46,7 +46,7 @@ export function isAzureOpenAiResourceEndpoint(endpoint: string | undefined): boo
  * True when the embedding base URL targets Azure (resource, APIM gateway, Cognitive Services, Foundry),
  * not public api.openai.com.
  */
-function isAzureOpenAiCompatibleEndpoint(endpoint: string | undefined): boolean {
+export function isAzureOpenAiCompatibleEndpoint(endpoint: string | undefined): boolean {
   if (typeof endpoint !== "string" || !endpoint.trim()) return false;
   // Use specific Azure AI/OpenAI domains only — `\.azure\.com` alone is too broad and would
   // match unrelated Azure services (portal.azure.com, devops.azure.com, etc.).

package/services/recall-pipeline.ts

@@ -24,6 +24,7 @@ import type { EmbeddingProvider } from "./embeddings.js";
 import { shouldSuppressEmbeddingError } from "./embeddings.js";
 import { expandQueryWithHyde } from "./hyde-helper.js";
 import { DEFAULT_INTERACTIVE_RECALL_POLICY, type InteractiveRecallPolicy } from "./retrieval-mode-policy.js";
+import { yieldEventLoop } from "../utils/event-loop-yield.js";
 
 async function embedWithAbortRace(
   embedPromise: Promise<number[]>,
@@ -130,6 +131,9 @@ export async function runRecallPipelineQuery(
   stageMs.fts = Date.now() - t0;
   sqliteResults = [...sqliteResults, ...ftsResults];
 
+  // FTS + lookup are synchronous SQLite — yield so gateway WebSocket/health can run (#931).
+  await yieldEventLoop();
+
   let lanceResults: SearchResult[] = [];
   const useSemantic = cfg.retrievalStrategies.includes("semantic");
 
@@ -232,6 +236,8 @@ export async function runRecallPipelineQuery(
     }
   }
 
+  await yieldEventLoop();
+
   t0 = Date.now();
   let results = mergeResults(sqliteResults, lanceResults, limitNum, factsDb);
   stageMs.merge = Date.now() - t0;

package/src/worker/narratives.ts

@@ -2,9 +2,9 @@ import type OpenAI from "openai";
 import type { EventLog } from "../../backends/event-log.js";
 import type { NarrativesDB } from "../../backends/narratives-db.js";
 import type { WorkflowStore } from "../../backends/workflow-store.js";
-import { chatCompleteWithRetry } from "../../services/chat.js";
+import { chatCompleteWithRetry, isAbortOrTransientLlmError } from "../../services/chat.js";
 import { capturePluginError } from "../../services/error-reporter.js";
-import { getSessionLogFileSuffix } from "../../utils/constants.js";
+import { getSessionLogFileSuffix, NARRATIVE_CHAT_TIMEOUT_MS } from "../../utils/constants.js";
 import { fillPrompt, loadPrompt } from "../../utils/prompt-loader.js";
 
 /** Session transcript basename for `sessionId` (suffix from OPENCLAW_SESSION_LOG_SUFFIX, default .jsonl). */
@@ -105,6 +105,7 @@ export async function buildDailyNarrative(params: BuildDailyNarrativeParams): Pr
       fallbackModels: fallbackModels ?? [],
       label: "memory-hybrid: narrative-summary",
       feature: "distill",
+      timeoutMs: NARRATIVE_CHAT_TIMEOUT_MS,
     });
     const normalized = normalizeNarrative(raw);
     if (!normalized || normalized === "NO_NARRATIVE") return false;
@@ -122,12 +123,20 @@ export async function buildDailyNarrative(params: BuildDailyNarrativeParams): Pr
     );
     return true;
   } catch (err) {
-    capturePluginError(err instanceof Error ? err : new Error(String(err)), {
-      subsystem: "narratives",
-      operation: "build-daily-narrative",
-      sessionId,
-    });
-    logger.warn(`memory-hybrid: narrative build failed for ${sessionId}: ${err}`);
+    const transient = isAbortOrTransientLlmError(err);
+    if (!transient) {
+      capturePluginError(err instanceof Error ? err : new Error(String(err)), {
+        subsystem: "narratives",
+        operation: "build-daily-narrative",
+        sessionId,
+      });
+    }
+    const detail = err instanceof Error ? err.message : String(err);
+    if (transient) {
+      logger.info?.(`memory-hybrid: narrative skipped (LLM unavailable or aborted) for ${sessionId}: ${detail}`);
+    } else {
+      logger.warn(`memory-hybrid: narrative build failed for ${sessionId}: ${err}`);
+    }
     return false;
   }
 }

package/tools/memory-tools.ts

@@ -19,7 +19,7 @@ import type { EventLog } from "../backends/event-log.js";
 import type { NarrativesDB } from "../backends/narratives-db.js";
 import { categoryToEventType } from "../backends/event-log.js";
 import type { EmbeddingProvider } from "../services/embeddings.js";
-import { AllEmbeddingProvidersFailed } from "../services/embeddings.js";
+import { AllEmbeddingProvidersFailed, shouldSuppressEmbeddingError } from "../services/embeddings.js";
 import type { EmbeddingRegistry } from "../services/embedding-registry.js";
 import { toFloat32Array } from "../services/embedding-registry.js";
 import type { PendingLLMWarnings } from "../services/chat.js";
@@ -1546,6 +1546,9 @@ export function registerMemoryTools(
               // Graceful degradation: store the fact without a vector.
               // The fact is still findable by structured/keyword search.
               api.logger.warn("memory-hybrid: Stored fact without embeddings — all providers unavailable");
+            } else if (shouldSuppressEmbeddingError(err)) {
+              // Ollama circuit breaker, 429, config errors, etc. — expected noise (#937); don't send to GlitchTip.
+              api.logger.warn(`memory-hybrid: embedding skipped (expected): ${err}`);
             } else {
               capturePluginError(err instanceof Error ? err : new Error(String(err)), {
                 subsystem: "embeddings",

package/utils/constants.ts

@@ -75,6 +75,9 @@ export const OLLAMA_COOLDOWN_MS = 5 * 60 * 1000; // 5 minutes
 /** Default timeout for chat completion requests (ms). */
 export const DEFAULT_CHAT_TIMEOUT_MS = 45_000;
 
+/** Daily narrative prompts aggregate many events — allow longer than default chat timeout (#935, #936). */
+export const NARRATIVE_CHAT_TIMEOUT_MS = 120_000;
+
 // VectorDB constants
 /** Threshold for warning about consecutive optimize failures. */
 export const VECTORDB_OPTIMIZE_FAILURE_WARN_THRESHOLD = 3;

package/utils/event-loop-yield.ts

@@ -0,0 +1,9 @@
+/**
+ * Yield to the Node.js event loop so I/O and WebSocket handlers (e.g. gateway health RPCs) can run.
+ * Use after heavy synchronous SQLite / merge work on the auto-recall path (#931).
+ */
+export async function yieldEventLoop(): Promise<void> {
+  await new Promise<void>((resolve) => {
+    setImmediate(resolve);
+  });
+}