@smyslenny/agent-memory 4.1.0-alpha.1 → 4.3.0

package/CHANGELOG.md

@@ -1,5 +1,43 @@
 # Changelog
 
+## 4.2.0 (2026-03-19)
+
+### 🛡️ Anti-Noise Hardening
+
+This release addresses the "heartbeat flood" problem where memory-sync cron
+ingested hundreds of low-value status observations (e.g. "HEARTBEAT_OK",
+"安静模式", "PR 无变化") into the curated memory store.
+
+#### Guard improvements
+- **Raised specificity threshold** for P2/P3 memories from 8 to 20 effective
+  characters
+- **CJK-aware length calculation**: CJK characters count as 3 effective chars
+  (reflecting their higher information density vs ASCII), preventing false
+  rejections of legitimate Chinese content
+
+#### Ingest noise filter
+- Added `isIngestNoise()` pre-filter in `extractIngestItems()` that skips lines
+  matching known noise patterns before they reach the Write Guard:
+  - Heartbeat status: `HEARTBEAT_OK`, `安静模式`, `不打扰`, `继续安静待命`
+  - Empty deltas: `无新 delta`, `无变化`, `无紧急`, `无新进展`
+  - System dumps: `openclaw status`, `openclaw gateway status`, `session_status`
+  - Stale PR observations: `PR #NNN 无变化`, `基线未变`, `轻量复查`
+  - Cron noise: `cron 会话`, `距上次心跳`, `危险区协议`
+
+#### Tidy expansion
+- `getDecayedMemories()` now includes **P2 (knowledge)** in cleanup candidates
+  (previously only P3 event). This means decayed low-vitality knowledge entries
+  will be cleaned up during sleep tidy, not just events.
+
+#### Govern env config
+- `maxMemories` can now be set via `AGENT_MEMORY_MAX_MEMORIES` environment
+  variable (default: 200)
+
+### ✅ Tests
+- 66 tests passing (19 files)
+- Added `tests/ingest/noise-filter.test.ts` covering heartbeat noise rejection,
+  meaningful content preservation, and mixed signal/noise handling
+
 ## 4.0.0-alpha.1 (2026-03-09)
 
 ### 🚀 Repositioning

package/README.md

@@ -22,7 +22,7 @@ AgentMemory is a SQLite-first memory layer for AI agents. It lets an agent:
 - **maintain** them over time with `reflect`, `reindex`, and feedback signals
 - **integrate** through **CLI**, **MCP stdio**, or **HTTP/SSE**
 
-Current release: **`4.0.0-alpha.1`**.
+Current release: **`4.3.0`**.
 
 Without an embedding provider, AgentMemory still works in **BM25-only mode**.
 With one configured, it adds **hybrid recall** and **semantic dedup**.

package/dist/bin/agent-memory.js

@@ -691,6 +691,44 @@ function createLocalHttpEmbeddingProvider(opts) {
     }
   };
 }
+function createGeminiEmbeddingProvider(opts) {
+  const id = stableProviderId(`gemini:${opts.model}`, `${opts.model}|${opts.dimension}`);
+  return {
+    id,
+    model: opts.model,
+    dimension: opts.dimension,
+    async embed(texts) {
+      if (texts.length === 0) return [];
+      const fetchFn = getFetch(opts.fetchImpl);
+      const url = `https://generativelanguage.googleapis.com/v1beta/models/${opts.model}:batchEmbedContents?key=${opts.apiKey}`;
+      const requests = texts.map((text) => ({
+        model: `models/${opts.model}`,
+        content: { parts: [{ text }] },
+        outputDimensionality: opts.dimension
+      }));
+      const response = await fetchFn(url, {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ requests })
+      });
+      if (!response.ok) {
+        const detail = await response.text().catch(() => "");
+        throw new Error(`Gemini embedding request failed: ${response.status} ${response.statusText}${detail ? ` \u2014 ${detail}` : ""}`);
+      }
+      const json2 = await response.json();
+      const embeddings = json2?.embeddings;
+      if (!Array.isArray(embeddings)) {
+        throw new Error("Gemini embedding response missing embeddings array");
+      }
+      return embeddings.map(
+        (entry, index) => assertEmbeddingVector(entry?.values, opts.dimension, `Gemini embedding item ${index}`)
+      );
+    },
+    async healthcheck() {
+      await this.embed(["healthcheck"]);
+    }
+  };
+}
 function normalizeEmbeddingBaseUrl(baseUrl) {
   return trimTrailingSlashes(baseUrl);
 }
@@ -708,7 +746,7 @@ function parseDimension(raw) {
 }
 function parseProvider(raw) {
   if (!raw) return null;
-  if (raw === "openai-compatible" || raw === "local-http") {
+  if (raw === "openai-compatible" || raw === "local-http" || raw === "gemini") {
     return raw;
   }
   throw new Error(`Unsupported embedding provider: ${raw}`);
@@ -719,8 +757,8 @@ function getEmbeddingProviderConfigFromEnv(env = process.env) {
   const baseUrl = env.AGENT_MEMORY_EMBEDDING_BASE_URL;
   const model = env.AGENT_MEMORY_EMBEDDING_MODEL;
   const dimension = parseDimension(env.AGENT_MEMORY_EMBEDDING_DIMENSION);
-  if (!baseUrl) {
-    throw new Error("AGENT_MEMORY_EMBEDDING_BASE_URL is required when embeddings are enabled");
+  if (!baseUrl && provider !== "gemini") {
+    throw new Error("AGENT_MEMORY_EMBEDDING_BASE_URL is required when embeddings are enabled (not needed for gemini provider)");
   }
   if (!model) {
     throw new Error("AGENT_MEMORY_EMBEDDING_MODEL is required when embeddings are enabled");
@@ -731,9 +769,12 @@ function getEmbeddingProviderConfigFromEnv(env = process.env) {
   if (provider === "openai-compatible" && !env.AGENT_MEMORY_EMBEDDING_API_KEY) {
     throw new Error("AGENT_MEMORY_EMBEDDING_API_KEY is required for openai-compatible providers");
   }
+  if (provider === "gemini" && !env.AGENT_MEMORY_EMBEDDING_API_KEY) {
+    throw new Error("AGENT_MEMORY_EMBEDDING_API_KEY is required for gemini provider (Google AI API key)");
+  }
   return {
     provider,
-    baseUrl,
+    baseUrl: baseUrl ?? "",
     model,
     dimension,
     apiKey: env.AGENT_MEMORY_EMBEDDING_API_KEY
@@ -742,8 +783,16 @@ function getEmbeddingProviderConfigFromEnv(env = process.env) {
 function createEmbeddingProvider(input, opts) {
   const normalized = {
     ...input,
-    baseUrl: normalizeEmbeddingBaseUrl(input.baseUrl)
+    baseUrl: input.baseUrl ? normalizeEmbeddingBaseUrl(input.baseUrl) : ""
   };
+  if (normalized.provider === "gemini") {
+    return createGeminiEmbeddingProvider({
+      model: normalized.model,
+      dimension: normalized.dimension,
+      apiKey: normalized.apiKey,
+      fetchImpl: opts?.fetchImpl
+    });
+  }
   if (normalized.provider === "openai-compatible") {
     return createOpenAICompatibleEmbeddingProvider({
       baseUrl: normalized.baseUrl,
@@ -770,13 +819,16 @@ function resolveEmbeddingProviderConfig(opts) {
   const model = opts?.config?.model ?? envConfig?.model;
   const dimension = opts?.config?.dimension ?? envConfig?.dimension;
   const apiKey = opts?.config?.apiKey ?? envConfig?.apiKey;
-  if (!provider || !baseUrl || !model || !dimension) {
+  if (!provider || !model || !dimension) {
     throw new Error("Incomplete embedding provider configuration");
   }
-  if (provider === "openai-compatible" && !apiKey) {
-    throw new Error("OpenAI-compatible embedding providers require an API key");
+  if (provider !== "gemini" && !baseUrl) {
+    throw new Error("baseUrl is required for non-gemini embedding providers");
+  }
+  if ((provider === "openai-compatible" || provider === "gemini") && !apiKey) {
+    throw new Error(`${provider} embedding provider requires an API key`);
   }
-  return { provider, baseUrl, model, dimension, apiKey };
+  return { provider, baseUrl: baseUrl ?? "", model, dimension, apiKey };
 }
 function getEmbeddingProvider(opts) {
   const config = resolveEmbeddingProviderConfig({ config: opts?.config, env: opts?.env });
@@ -2187,9 +2239,11 @@ function fourCriterionGate(input) {
   const content = input.content.trim();
   const failed = [];
   const priority = input.priority ?? (input.type === "identity" ? 0 : input.type === "emotion" ? 1 : input.type === "knowledge" ? 2 : 3);
-  const minLength = priority <= 1 ? 4 : 8;
-  const specificity = content.length >= minLength ? Math.min(1, content.length / 50) : 0;
-  if (specificity === 0) failed.push(`specificity (too short: ${content.length} < ${minLength} chars)`);
+  const cjkCount = (content.match(/[\u4e00-\u9fff\u3400-\u4dbf]/g) ?? []).length;
+  const effectiveLength = content.length + cjkCount * 2;
+  const minLength = priority <= 1 ? 4 : 20;
+  const specificity = effectiveLength >= minLength ? Math.min(1, effectiveLength / 50) : 0;
+  if (specificity === 0) failed.push(`specificity (too short: effective ${effectiveLength} < ${minLength} chars)`);
   const tokens = tokenize(content);
   const novelty = tokens.length >= 1 ? Math.min(1, tokens.length / 5) : 0;
   if (novelty === 0) failed.push("novelty (no meaningful tokens after filtering)");
@@ -2420,9 +2474,9 @@ function getDecayedMemories(db, threshold = 0.05, opts) {
   const agentId = opts?.agent_id;
   return db.prepare(
     agentId ? `SELECT id, content, vitality, priority FROM memories
-           WHERE vitality < ? AND priority >= 3 AND agent_id = ?
+           WHERE vitality < ? AND priority >= 2 AND agent_id = ?
            ORDER BY vitality ASC` : `SELECT id, content, vitality, priority FROM memories
-           WHERE vitality < ? AND priority >= 3
+           WHERE vitality < ? AND priority >= 2
            ORDER BY vitality ASC`
   ).all(...agentId ? [threshold, agentId] : [threshold]);
 }
@@ -2520,7 +2574,8 @@ function rankEvictionCandidates(db, opts) {
 }
 function runGovern(db, opts) {
   const agentId = opts?.agent_id;
-  const maxMemories = opts?.maxMemories ?? 200;
+  const envMax = Number.parseInt(process.env.AGENT_MEMORY_MAX_MEMORIES ?? "", 10);
+  const maxMemories = opts?.maxMemories ?? (Number.isFinite(envMax) && envMax > 0 ? envMax : 200);
   let orphanPaths = 0;
   let emptyMemories = 0;
   let evicted = 0;