@modusensus/dsh-mneme 0.5.0 → 0.5.2

package/lib/store.js

@@ -12,6 +12,7 @@ CREATE TABLE IF NOT EXISTS memories (
   forgotten   INTEGER NOT NULL DEFAULT 0,
   archived    INTEGER NOT NULL DEFAULT 0,
   source      TEXT,
+  session_id  TEXT,
   content_history TEXT,
   embedding   TEXT,
   epistemic_status TEXT NOT NULL DEFAULT 'subjective',
@@ -321,6 +322,7 @@ function toRow(row) {
     forgotten: row.forgotten === 1,
     archived: row.archived === 1,
     source: row.source ?? undefined,
+    session_id: row.session_id ?? undefined,
     content_history: parseJsonArray(row.content_history),
     quality_score: row.quality_score !== null && row.quality_score !== undefined ? Number(row.quality_score) : undefined,
     epistemic_status: row.epistemic_status ?? "subjective",
@@ -571,6 +573,7 @@ export function createStore(path) {
   addColumn("memories", "epistemic_status", "ALTER TABLE memories ADD COLUMN epistemic_status TEXT NOT NULL DEFAULT 'subjective'");
   addColumn("memories", "content_history", "ALTER TABLE memories ADD COLUMN content_history TEXT");
   addColumn("memories", "quality_score", "ALTER TABLE memories ADD COLUMN quality_score REAL");
+  addColumn("memories", "session_id", "ALTER TABLE memories ADD COLUMN session_id TEXT");
 
   // Legacy dream_runs without policy_epoch → backfill with the default epoch.
   addColumn("dream_runs", "policy_epoch", "ALTER TABLE dream_runs ADD COLUMN policy_epoch INTEGER NOT NULL DEFAULT 0");
@@ -657,8 +660,8 @@ export function createStore(path) {
       : inferEpistemicStatus(memory);
     runAtomically(() => {
       db.prepare(
-        `INSERT INTO memories (id, type, title, content, tags, importance, forgotten, archived, source, content_history, quality_score, embedding, epistemic_status, created_at, updated_at)
-         VALUES (?, ?, ?, ?, ?, ?, 0, ?, ?, ?, ?, ?, ?, ?, ?)`
+        `INSERT INTO memories (id, type, title, content, tags, importance, forgotten, archived, source, session_id, content_history, quality_score, embedding, epistemic_status, created_at, updated_at)
+         VALUES (?, ?, ?, ?, ?, ?, 0, ?, ?, ?, ?, ?, ?, ?, ?, ?)`
       ).run(
         id,
         type,
@@ -668,6 +671,7 @@ export function createStore(path) {
         importance,
         memory.archived ? 1 : 0,
         memory.source ?? null,
+        memory.session_id ?? null,
         JSON.stringify(memory.content_history ?? []),
         Number.isFinite(memory.quality_score) ? memory.quality_score : null,
         embedding,

package/lib/summarize.js

@@ -182,7 +182,11 @@ export function createSummarizer(ctx, service, config) {
         .join("");
       const entries = parseSummaryJson(text || assembledText);
       for (const entry of entries) {
-        service.saveWithDedupe({ ...entry, source: `session:${session.id}` });
+        // Provenance: the summarizer runs on a real session (turn/end hook), so
+        // session.id is always available here — it rides both the human-readable
+        // source label and the structured session_id column (v0.5.x memory
+        // provenance, the raw material for v0.6.0 reasoning-path / drift analysis).
+        service.saveWithDedupe({ ...entry, source: `session:${session.id}`, session_id: session.id });
       }
     } finally {
       if (audit) {

package/lib/tools.js

@@ -16,6 +16,7 @@ const MEMORY_ITEM_SCHEMA = {
     tags: { type: "array", items: { type: "string" } },
     importance: { type: "integer", required: true },
     source: { type: "string" },
+    session_id: { type: "string" },
     created_at: { type: "string", required: true },
     updated_at: { type: "string", required: true }
   }
@@ -48,14 +49,18 @@ export function createTools(ctx, service, config, embedder) {
         },
         render: (_args, value) => TEXT_OUTPUT(`memory ${value.action}: ${value.id}`)
       },
-      async execute(args) {
+      async execute(args, exec) {
         const { action, memory } = service.saveWithDedupe({
           type: args.type,
           title: args.title,
           content: args.content,
           tags: args.tags ?? [],
           importance: args.importance ?? 3,
-          source: args.source ?? "tool"
+          source: args.source ?? "tool",
+          // Provenance: the session that issued the tool call. exec.agent is
+          // set by the agent loop (undefined in unit tests / direct calls) —
+          // absent a session, session_id stays null rather than fabricating one.
+          session_id: exec?.agent?.session?.id ?? undefined
         });
         return { action, id: memory.id };
       }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@modusensus/dsh-mneme",
   "description": "Cross-session memory plugin for DeepSeek Harness with autoDream consolidation: SQLite store, Markdown mirrors, 7 model tools, automatic injection, session summarization, user profile/rules, custom slash commands, vector (semantic) search, and a Web GUI panel",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "license": "MIT",
   "repository": {
     "type": "git",

package/scripts/sync-lib.js

@@ -5,7 +5,12 @@
 //
 // Usage: npm run sync   (also run automatically by `npm pack`/`npm publish`
 // via the prepack hook, so a published tarball always ships a fresh lib/).
-import { cpSync, existsSync, mkdirSync, readdirSync } from "node:fs";
+//
+// Note: copyFileSync (not cpSync) is used on purpose — cpSync removes the
+// destination first, which fails with EPERM/unlink on Windows when the path
+// is long enough to trigger the \\?\ extended-prefix (observed on publish).
+// copyFileSync truncates and rewrites in place, so it survives long paths.
+import { copyFileSync, existsSync, mkdirSync, readdirSync } from "node:fs";
 import { join, relative } from "node:path";
 import { fileURLToPath } from "node:url";
 
@@ -31,7 +36,7 @@ for (const file of walk(srcDir)) {
   const rel = relative(srcDir, file);
   const dest = join(libDir, rel);
   mkdirSync(join(dest, ".."), { recursive: true });
-  cpSync(file, dest);
+  copyFileSync(file, dest);
   copied++;
   console.log(`synced  ${rel}`);
 }

package/src/hot-memory.js

@@ -1,46 +1,53 @@
-// Session-scoped hot memory (v0.5.0 召回率优化 1.3): a short-term buffer of
-// the latest dialogue rounds, kept strictly apart from the long-term memory
-// store. The injector renders it ahead of the long-term recall block so the
-// agent sees "what we were just talking about" without those rounds ever
-// being persisted as memories. Bounded two ways: maxRounds (count) and
-// maxTokens (budget) — whichever evicts first.
-
-// CJK-aware token estimate: one Chinese character ≈ 0.6 tokens (clustering
-// behavior of mainstream tokenizers), one ASCII char ≈ 0.25.
-export function estimateTokens(text) {
-  const s = String(text ?? "");
-  let cjk = 0;
-  for (const ch of s) if (ch >= "\u4e00" && ch <= "\u9fff") cjk++;
-  return Math.ceil(cjk * 0.6 + (s.length - cjk) * 0.25);
-}
-
-/**
- * @param {{maxRounds?: number, maxTokens?: number}} opts
- * @returns {{add(round: {query: string, response?: string}): void,
- *            getContext(): string,
- *            rounds(): Array, clear(): void}}
- */
-export function createHotMemory({ maxRounds = 5, maxTokens = 2000 } = {}) {
-  const buffer = [];
-
-  function totalTokens() {
-    return buffer.reduce(
-      (sum, r) => sum + estimateTokens(`Q: ${r.query}\nA: ${r.response ?? ""}`),
-      0
-    );
-  }
-
-  return {
-    add(round) {
-      if (!round?.query) return;
-      buffer.push({ query: String(round.query), response: String(round.response ?? "") });
-      while (buffer.length > maxRounds) buffer.shift();
-      while (buffer.length > 1 && totalTokens() > maxTokens) buffer.shift();
-    },
-    getContext() {
-      return buffer.map((r) => `Q: ${r.query}\nA: ${r.response ?? ""}`).join("\n\n");
-    },
-    rounds: () => [...buffer],
-    clear() { buffer.length = 0; }
-  };
-}
+// Session-scoped hot memory (v0.5.0 召回率优化 1.3): a short-term buffer of
+// the latest dialogue rounds, kept strictly apart from the long-term memory
+// store. The injector renders it ahead of the long-term recall block so the
+// agent sees "what we were just talking about" without those rounds ever
+// being persisted as memories. Bounded two ways: maxRounds (count) and
+// maxTokens (budget) — whichever evicts first.
+
+// CJK-aware token estimate: one Chinese character ≈ 0.6 tokens (clustering
+// behavior of mainstream tokenizers), one ASCII char ≈ 0.25.
+export function estimateTokens(text) {
+  const s = String(text ?? "");
+  let cjk = 0;
+  for (const ch of s) if (ch >= "\u4e00" && ch <= "\u9fff") cjk++;
+  return Math.ceil(cjk * 0.6 + (s.length - cjk) * 0.25);
+}
+
+/**
+ * @param {{maxRounds?: number, maxTokens?: number}} opts
+ * @returns {{add(round: {query: string, response?: string}): void,
+ *            getContext(): string,
+ *            rounds(): Array, clear(): void}}
+ */
+export function createHotMemory({ maxRounds = 5, maxTokens = 2000 } = {}) {
+  // Entry defense: a non-positive or non-integer maxRounds (0, -1, 1.5, NaN,
+  // null, "2") would make the eviction while-loop unbounded — the buffer can
+  // never shrink below `buffer.length > maxRounds`, so `add` would spin forever.
+  // Fall back to the defaults so a hostile/buggy caller can never wedge the
+  // hot-memory buffer in an infinite loop.
+  maxRounds = (Number.isInteger(maxRounds) && maxRounds > 0) ? maxRounds : 5;
+  maxTokens = (Number.isFinite(maxTokens) && maxTokens > 0) ? maxTokens : 2000;
+  const buffer = [];
+
+  function totalTokens() {
+    return buffer.reduce(
+      (sum, r) => sum + estimateTokens(`Q: ${r.query}\nA: ${r.response ?? ""}`),
+      0
+    );
+  }
+
+  return {
+    add(round) {
+      if (!round?.query) return;
+      buffer.push({ query: String(round.query), response: String(round.response ?? "") });
+      while (buffer.length > maxRounds) buffer.shift();
+      while (buffer.length > 1 && totalTokens() > maxTokens) buffer.shift();
+    },
+    getContext() {
+      return buffer.map((r) => `Q: ${r.query}\nA: ${r.response ?? ""}`).join("\n\n");
+    },
+    rounds: () => [...buffer],
+    clear() { buffer.length = 0; }
+  };
+}