openclaw-amem 2.0.1 → 2.1.0

package/README.md

@@ -89,10 +89,11 @@ Add `openclaw-amem` to your allowed plugins and hook it into the `memory` slot:
 openclaw gateway restart
 ```
 
-First run downloads the embedding model (`bge-m3`, 1.08 GB) and caches it. Later
-restarts are instant. Want something smaller? `AMEM_EMBED_MODEL=onnx-community/gte-multilingual-base`
-is about a third the size and reads text just as long — set it **before** you have
-memories, since changing it afterwards means a [migration](https://amem.owo.lc/reference/embedding-models#changing-the-model-on-a-store-you-already-have).
+First run downloads the embedding model (`bge-m3`, 2.27 GB) and caches it. Later
+restarts are instant. If that is too much, the only smaller model worth setting is
+`AMEM_EMBED_MODEL=Xenova/bge-small-zh-v1.5` — 25 MB, Chinese only, and it caps at
+512 tokens. Set it **before** you have memories, since changing it afterwards means
+a [migration](https://amem.owo.lc/reference/embedding-models#changing-the-model-on-a-store-you-already-have).
 
 Upgrading from 1.x downloads nothing. Your existing memories keep the model that
 built them, and the plugin says so at startup along with the one command that

package/dist/index.js

@@ -76,9 +76,28 @@ function getEmbeddingDevice() {
   return process.env.AMEM_EMBED_DEVICE?.trim() || void 0;
 }
 function getEmbeddingDtype() {
-  const explicit = process.env.AMEM_EMBED_DTYPE?.trim();
-  if (explicit) return explicit;
-  return getEmbeddingModel() === DEFAULT_EMBEDDING_MODEL ? DEFAULT_MODEL_DTYPE : void 0;
+  return process.env.AMEM_EMBED_DTYPE?.trim() || void 0;
+}
+function applyModelPaths(env) {
+  const cache = process.env.AMEM_MODEL_CACHE?.trim();
+  if (cache) env.cacheDir = cache;
+  const local = process.env.AMEM_MODEL_DIR?.trim();
+  if (local) {
+    env.localModelPath = local;
+    env.allowLocalModels = true;
+  }
+}
+function makeProgressReporter() {
+  const lastPct = /* @__PURE__ */ new Map();
+  return (e) => {
+    if (e.status !== "progress" || !e.file || typeof e.progress !== "number") return;
+    if (!e.file.endsWith(".onnx") && !e.file.endsWith(".onnx_data")) return;
+    const pct = Math.floor(e.progress / 10) * 10;
+    if (lastPct.get(e.file) === pct) return;
+    lastPct.set(e.file, pct);
+    const size = e.total ? ` of ${(e.total / 1e9).toFixed(2)} GB` : "";
+    console.log(`[amem] downloading ${e.file}: ${pct}%${size}`);
+  };
 }
 function extractorKey() {
   return `${getEmbeddingModel()}|${getEmbeddingDevice() ?? ""}|${getEmbeddingDtype() ?? ""}`;
@@ -89,11 +108,13 @@ async function getExtractor() {
   if (!pipeline) {
     const mod = await import("@huggingface/transformers");
     pipeline = mod.pipeline;
+    applyModelPaths(mod.env);
   }
   const device = getEmbeddingDevice();
   const dtype = getEmbeddingDtype();
   extractor = await pipeline("feature-extraction", getEmbeddingModel(), {
     revision: "main",
+    progress_callback: makeProgressReporter(),
     // Omitted entirely when unset, so an unconfigured install gets exactly the
     // library defaults it got before these existed.
     ...device ? { device } : {},
@@ -167,7 +188,7 @@ function cosineSimilarity(a, b) {
   for (let i = 0; i < a.length; i++) dot += a[i] * b[i];
   return dot;
 }
-var pipeline, extractor, loadedKey, cachedDim, DEFAULT_EMBEDDING_MODEL, LEGACY_DEFAULT_EMBEDDING_MODEL, LEGACY_DEFAULT_DIM, DEFAULT_MODEL_DTYPE, pinnedModel, CLS_POOLED_MODELS;
+var pipeline, extractor, loadedKey, cachedDim, DEFAULT_EMBEDDING_MODEL, LEGACY_DEFAULT_EMBEDDING_MODEL, LEGACY_DEFAULT_DIM, pinnedModel, CLS_POOLED_MODELS;
 var init_embedding = __esm({
   "../amem-core/src/embedding.ts"() {
     "use strict";
@@ -178,7 +199,6 @@ var init_embedding = __esm({
     DEFAULT_EMBEDDING_MODEL = "Xenova/bge-m3";
     LEGACY_DEFAULT_EMBEDDING_MODEL = "Xenova/paraphrase-multilingual-MiniLM-L12-v2";
     LEGACY_DEFAULT_DIM = 384;
-    DEFAULT_MODEL_DTYPE = "fp16";
     pinnedModel = null;
     CLS_POOLED_MODELS = /* @__PURE__ */ new Set([
       "bge-m3",
@@ -377,6 +397,10 @@ async function scrollIdsRaw(collection, limit = 1e4) {
 async function deleteCollectionRaw(collection) {
   await qdrant("DELETE", `/collections/${collection}`);
 }
+async function snapshotCollectionRaw(collection) {
+  const r = await qdrant("POST", `/collections/${collection}/snapshots`);
+  return { name: r.name, size: r.size ?? 0 };
+}
 async function resolveAliasRaw(alias) {
   try {
     const res = await qdrant("GET", `/aliases`);
@@ -2559,6 +2583,7 @@ async function switchToMigrated(opts) {
   const { name, to } = opts;
   const log = opts.logger?.info ?? ((m) => console.log(m));
   if (name === to) throw new Error(`switch: "${name}" and "${to}" are the same collection`);
+  let snap;
   const already = await resolveAliasRaw(name);
   if (already === to) {
     log(`[switch] "${name}" already points at "${to}" \u2014 nothing to do`);
@@ -2574,6 +2599,10 @@ async function switchToMigrated(opts) {
       );
     }
     log(`[switch] verified ${targetCount} in "${to}" against ${sourceCount} in "${name}"`);
+    if (opts.snapshot !== false) {
+      snap = await snapshotCollectionRaw(name);
+      log(`[switch] snapshotted "${name}" \u2192 ${snap.name} (${(snap.size / 1e6).toFixed(0)} MB)`);
+    }
     await deleteCollectionRaw(name);
     log(`[switch] dropped "${name}"`);
     await createAliasRaw(name, to);
@@ -2581,7 +2610,7 @@ async function switchToMigrated(opts) {
     await setAliasRaw(name, to);
   }
   log(`[switch] "${name}" now resolves to "${to}"`);
-  return { name, to, moved: targetCount };
+  return { name, to, moved: targetCount, snapshot: snap };
 }
 var init_migrate = __esm({
   "../amem-core/src/migrate.ts"() {

package/openclaw.plugin.json

@@ -3,7 +3,7 @@
   "name": "amem",
   "description": "Catches memories that contradict each other. Notes rewrite themselves as new ones arrive, link into a graph, and stay separated per agent and per person. Runs on a small model. Requires a local Qdrant. 中文走 jieba 分词，提示词有中文版。",
   "icon": "https://amem.owo.lc/logo.png",
-  "version": "2.0.1",
+  "version": "2.1.0",
   "kind": "memory",
   "contracts": {
     "tools": [
@@ -35,6 +35,8 @@
           "AMEM_EMBED_POOLING",
           "AMEM_EMBED_DEVICE",
           "AMEM_EMBED_DTYPE",
+          "AMEM_MODEL_CACHE",
+          "AMEM_MODEL_DIR",
           "AMEM_COLLECTION",
           "AMEM_DATA_DIR",
           "AMEM_EVO_COUNTER_PATH",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openclaw-amem",
-  "version": "2.0.1",
+  "version": "2.1.0",
   "description": "Catches memories that contradict each other. Notes rewrite themselves as new ones arrive, link into a graph, and stay separated per agent and per person. Runs on a small model. Requires a local Qdrant. 中文走 jieba 分词，提示词有中文版。",
   "license": "MIT",
   "main": "dist/index.js",