memor-cli 0.5.0__tar.gz → 0.6.0__tar.gz

{memor_cli-0.5.0/memor_cli.egg-info → memor_cli-0.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memor-cli
-Version: 0.5.0
+Version: 0.6.0
 Summary: Measured memory for coding agents. Fire and forget — no API keys needed.
 Author-email: Nimit Bhandari <nimitbhandari17@gmail.com>
 License-Expression: MIT
@@ -46,7 +46,7 @@ Dynamic: license-file
 ```
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-227%20passing-brightgreen.svg)]()
+[![Tests](https://img.shields.io/badge/tests-256%20passing-brightgreen.svg)]()
 [![Python](https://img.shields.io/badge/python-3.11%2B-blue.svg)]()
 [![PyPI](https://img.shields.io/pypi/v/memor-cli.svg)](https://pypi.org/project/memor-cli/)
 
@@ -113,7 +113,7 @@ memor dashboard
 
 **Two background processes:**
 
-1. **Daemon** — polls `~/.claude/projects/` for transcripts, embeds chunks, runs distillation, analyzes feedback, compacts duplicates, tracks session-level token usage. All local.
+1. **Daemon** — polls `~/.claude/projects/` for transcripts, embeds chunks, runs distillation, analyzes feedback (positive and negative), promotes cross-project patterns to global scope, compacts duplicates, tracks session-level token usage. All local.
 2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms.
 
 **No API keys required.** Embeddings run locally via [model2vec](https://github.com/MinishLab/model2vec) (potion-base-8M, 256-dim). Vectors stored in [sqlite-vec](https://github.com/asg017/sqlite-vec). Everything runs on your machine.
@@ -146,7 +146,12 @@ This means a relevant decision from yesterday beats a vaguely-related chunk from
 
 ### Feedback Loop
 
-Memor tracks whether recalled memories actually get used by the agent. After each session, the daemon analyzes the transcript to detect if recalled content appeared in the agent's responses. Memories that consistently prove useful get quality boosts; memories never recalled in 30+ days get automatically deactivated. Near-duplicate memories are compacted into one.
+Memor tracks whether recalled memories actually get used by the agent — and whether they actively hurt. After each session, the daemon analyzes the transcript in both directions:
+
+- **Positive signal** — n-gram overlap or semantic similarity between recalled content and the agent's response. Memories that consistently prove useful get quality boosts.
+- **Negative signal** — user rejection ("no that's wrong", "we switched to X") or assistant contradiction ("however, looking at the current code, we actually use Y"). Memories that get corrected receive a quality penalty, making them less likely to be recalled next time.
+
+The quality formula is Bayesian: `(uses - negatives + 1) / (recalls + 2)`. One correction weighs as much as one positive use, so harmful memories drop fast. Memories never recalled in 30+ days get automatically deactivated. Near-duplicate memories are compacted into one.
 
 ---
 
@@ -161,6 +166,20 @@ Memories are automatically classified as `decision`, `bugfix`, `lesson`, `snippe
 
 ---
 
+## Global Memories
+
+Some patterns aren't project-specific — they're yours. "Always use type hints." "Structure FastAPI apps with a `routes/` directory." "Prefer composition over inheritance."
+
+Memor detects these automatically. When the same pattern appears in **3 or more projects** (measured by embedding similarity), the daemon promotes it to a `_global` scope:
+
+- **Global memories are recalled everywhere** — they show up in every project's search results alongside project-specific memories.
+- **Source duplicates are deactivated** — the per-project copies get superseded by the single global version, reducing clutter.
+- **No manual tagging** — promotion is fully automatic, based on cross-project clustering.
+
+This means your coding habits and preferences follow you into new projects from the first prompt, without you having to re-explain anything.
+
+---
+
 ## Dashboard
 
 ```bash
@@ -217,7 +236,8 @@ memor/
 +-- project.py            Git-root project resolver (filesystem-aware)
 +-- recall.py             Shared recall core (used by hook + skill)
 +-- redact.py             Secret detection and redaction at ingest
-+-- feedback.py           Implicit feedback analyzer (usage detection)
++-- feedback.py           Feedback analyzer (positive usage + negative signals)
++-- global_memories.py    Cross-project promotion to _global scope
 |
 +-- retrieve/
 |   +-- retriever.py      Hybrid retrieval (dense + BM25, RRF) + relevance gate + scoring
@@ -291,7 +311,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 215 tests
+pytest  # 251 tests
 ```
 
 ---

{memor_cli-0.5.0 → memor_cli-0.6.0}/README.md

@@ -9,7 +9,7 @@
 ```
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-227%20passing-brightgreen.svg)]()
+[![Tests](https://img.shields.io/badge/tests-256%20passing-brightgreen.svg)]()
 [![Python](https://img.shields.io/badge/python-3.11%2B-blue.svg)]()
 [![PyPI](https://img.shields.io/pypi/v/memor-cli.svg)](https://pypi.org/project/memor-cli/)
 
@@ -76,7 +76,7 @@ memor dashboard
 
 **Two background processes:**
 
-1. **Daemon** — polls `~/.claude/projects/` for transcripts, embeds chunks, runs distillation, analyzes feedback, compacts duplicates, tracks session-level token usage. All local.
+1. **Daemon** — polls `~/.claude/projects/` for transcripts, embeds chunks, runs distillation, analyzes feedback (positive and negative), promotes cross-project patterns to global scope, compacts duplicates, tracks session-level token usage. All local.
 2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms.
 
 **No API keys required.** Embeddings run locally via [model2vec](https://github.com/MinishLab/model2vec) (potion-base-8M, 256-dim). Vectors stored in [sqlite-vec](https://github.com/asg017/sqlite-vec). Everything runs on your machine.
@@ -109,7 +109,12 @@ This means a relevant decision from yesterday beats a vaguely-related chunk from
 
 ### Feedback Loop
 
-Memor tracks whether recalled memories actually get used by the agent. After each session, the daemon analyzes the transcript to detect if recalled content appeared in the agent's responses. Memories that consistently prove useful get quality boosts; memories never recalled in 30+ days get automatically deactivated. Near-duplicate memories are compacted into one.
+Memor tracks whether recalled memories actually get used by the agent — and whether they actively hurt. After each session, the daemon analyzes the transcript in both directions:
+
+- **Positive signal** — n-gram overlap or semantic similarity between recalled content and the agent's response. Memories that consistently prove useful get quality boosts.
+- **Negative signal** — user rejection ("no that's wrong", "we switched to X") or assistant contradiction ("however, looking at the current code, we actually use Y"). Memories that get corrected receive a quality penalty, making them less likely to be recalled next time.
+
+The quality formula is Bayesian: `(uses - negatives + 1) / (recalls + 2)`. One correction weighs as much as one positive use, so harmful memories drop fast. Memories never recalled in 30+ days get automatically deactivated. Near-duplicate memories are compacted into one.
 
 ---
 
@@ -124,6 +129,20 @@ Memories are automatically classified as `decision`, `bugfix`, `lesson`, `snippe
 
 ---
 
+## Global Memories
+
+Some patterns aren't project-specific — they're yours. "Always use type hints." "Structure FastAPI apps with a `routes/` directory." "Prefer composition over inheritance."
+
+Memor detects these automatically. When the same pattern appears in **3 or more projects** (measured by embedding similarity), the daemon promotes it to a `_global` scope:
+
+- **Global memories are recalled everywhere** — they show up in every project's search results alongside project-specific memories.
+- **Source duplicates are deactivated** — the per-project copies get superseded by the single global version, reducing clutter.
+- **No manual tagging** — promotion is fully automatic, based on cross-project clustering.
+
+This means your coding habits and preferences follow you into new projects from the first prompt, without you having to re-explain anything.
+
+---
+
 ## Dashboard
 
 ```bash
@@ -180,7 +199,8 @@ memor/
 +-- project.py            Git-root project resolver (filesystem-aware)
 +-- recall.py             Shared recall core (used by hook + skill)
 +-- redact.py             Secret detection and redaction at ingest
-+-- feedback.py           Implicit feedback analyzer (usage detection)
++-- feedback.py           Feedback analyzer (positive usage + negative signals)
++-- global_memories.py    Cross-project promotion to _global scope
 |
 +-- retrieve/
 |   +-- retriever.py      Hybrid retrieval (dense + BM25, RRF) + relevance gate + scoring
@@ -254,7 +274,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 215 tests
+pytest  # 251 tests
 ```
 
 ---

memor_cli-0.6.0/memor/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.6.0"

{memor_cli-0.5.0 → memor_cli-0.6.0}/memor/daemon.py

@@ -267,6 +267,16 @@ def run_poll_cycle(
         except Exception:
             pass
 
+    # Promote cross-project patterns to global scope
+    if new_ingested:
+        try:
+            from memor.global_memories import run_promotion
+            promoted = run_promotion(store, embedder, min_projects=3)
+            if promoted > 0:
+                print(f"  promoted {promoted} memories to global scope")
+        except Exception:
+            pass
+
     # Compact near-duplicate memories (run occasionally, not every cycle)
     if new_ingested:
         try:

{memor_cli-0.5.0 → memor_cli-0.6.0}/memor/dashboard/server.py

@@ -27,6 +27,7 @@ def create_app(db_path: str | None = None) -> FastAPI:
 
     @app.get("/api/summary")
     def summary():
+        from memor.types import GLOBAL_PROJECT
         store = _store()
         recall_stats = store.get_recall_stats()
         ingestion = {}
@@ -38,17 +39,22 @@ def create_app(db_path: str | None = None) -> FastAPI:
         project_count = store.db.execute(
             "SELECT COUNT(DISTINCT project) as c FROM artifacts"
         ).fetchone()["c"]
+        global_count = store.db.execute(
+            "SELECT COUNT(*) as c FROM artifacts WHERE active=1 AND project=? AND kind='memory'",
+            (GLOBAL_PROJECT,)
+        ).fetchone()["c"]
         recall_stats["ingestion"] = ingestion
         recall_stats["project_count"] = project_count
+        recall_stats["global_memories"] = global_count
         return recall_stats
 
     @app.get("/api/projects")
     def projects():
         store = _store()
         recall_stats = store.get_project_stats()
-        if recall_stats:
-            return recall_stats
-        rows = store.db.execute("""
+        recall_projects = {r["project"] for r in recall_stats} if recall_stats else set()
+
+        artifact_rows = store.db.execute("""
             SELECT project,
                    COUNT(*) as artifacts,
                    SUM(CASE WHEN kind='session_chunk' THEN 1 ELSE 0 END) as chunks,
@@ -59,7 +65,15 @@ def create_app(db_path: str | None = None) -> FastAPI:
             GROUP BY project
             ORDER BY artifacts DESC
         """).fetchall()
-        return [dict(r) for r in rows]
+
+        if not recall_stats:
+            return [dict(r) for r in artifact_rows]
+
+        result = list(recall_stats)
+        for r in artifact_rows:
+            if r["project"] not in recall_projects:
+                result.append(dict(r))
+        return result
 
     @app.get("/api/recalls")
     def recalls(limit: int = Query(50, ge=1, le=500),
@@ -71,7 +85,9 @@ def create_app(db_path: str | None = None) -> FastAPI:
     def quality():
         store = _store()
         rows = store.db.execute("""
-            SELECT q.artifact_id, q.recall_count, q.use_count, q.quality_score,
+            SELECT q.artifact_id, q.recall_count, q.use_count,
+                   COALESCE(q.negative_count, 0) as negative_count,
+                   q.quality_score,
                    a.project, a.kind, json_extract(a.meta, '$.mem_type') as mem_type,
                    substr(a.text, 1, 100) as preview
             FROM memory_quality q
@@ -152,6 +168,11 @@ def create_app(db_path: str | None = None) -> FastAPI:
             return {"status": "no_runs"}
         return result
 
+    @app.get("/api/version")
+    def version():
+        from memor import __version__
+        return {"version": __version__}
+
     @app.get("/api/health")
     def health():
         store = _store()

{memor_cli-0.5.0 → memor_cli-0.6.0}/memor/dashboard/static/index.html

@@ -113,7 +113,7 @@
   /* ── Metric cards ──────────────────────────────────────── */
   .metric-grid {
     display: grid;
-    grid-template-columns: repeat(4, 1fr);
+    grid-template-columns: repeat(5, 1fr);
     gap: 14px;
   }
   @media (max-width: 900px) { .metric-grid { grid-template-columns: repeat(2, 1fr); } }
@@ -300,6 +300,16 @@
   .badge-extractive::before { background: var(--warn); }
   .badge-no_hits { background: var(--neutral-dim); color: var(--neutral); }
   .badge-no_hits::before { background: var(--neutral); }
+  .badge-global { background: rgba(120,100,255,0.12); color: #a090ff; }
+  .badge-global::before { background: #a090ff; }
+  .badge-negative { background: rgba(232,84,84,0.12); color: var(--error); }
+  .badge-negative::before { background: var(--error); }
+
+  .quality-bar { display: flex; height: 6px; border-radius: 3px; overflow: hidden; background: var(--surface3); width: 100%; min-width: 60px; }
+  .quality-fill { height: 100%; border-radius: 3px; transition: width 0.3s; }
+  .quality-fill-ok { background: var(--ok); }
+  .quality-fill-warn { background: var(--warn); }
+  .quality-fill-error { background: var(--error); }
 
   .empty-state {
     text-align: center; padding: 40px 20px;
@@ -319,6 +329,7 @@
   <div class="topbar-inner">
     <div class="breadcrumb">Dashboard &rsaquo; <span>Overview</span></div>
     <div class="topbar-right">
+      <span id="version-badge" class="mono" style="font-size:11px;color:var(--text-muted);margin-right:12px;"></span>
       <div class="topbar-time">
         <span class="refresh-dot"></span>
         <span id="last-updated">Live</span>
@@ -369,6 +380,13 @@
         </div>
         <div class="mini-bars" id="mb-projects"></div>
       </div>
+      <div class="metric-card">
+        <div class="metric-info">
+          <div class="metric-label">Global</div>
+          <div class="metric-value" id="m-global">&ndash;</div>
+          <div class="metric-sub">cross-project memories</div>
+        </div>
+      </div>
       <div class="metric-card">
         <div class="metric-info">
           <div class="metric-label">Total Recalls</div>
@@ -485,6 +503,30 @@
     </div>
   </section>
 
+  <!-- ── Memory Quality ──────────────────────────────────── -->
+  <section>
+    <div class="table-card" id="quality-section" style="display:none;">
+      <div class="table-header">
+        <div class="table-title"><span class="dot"></span> Memory Quality</div>
+      </div>
+      <table>
+        <thead>
+          <tr>
+            <th>Memory</th>
+            <th>Project</th>
+            <th>Type</th>
+            <th>Recalls</th>
+            <th>Used</th>
+            <th>Rejected</th>
+            <th>Score</th>
+            <th style="min-width:80px;"></th>
+          </tr>
+        </thead>
+        <tbody id="quality-body"></tbody>
+      </table>
+    </div>
+  </section>
+
   <!-- ── Recent recalls ─────────────────────────────────── -->
   <section>
     <div class="table-card">
@@ -603,6 +645,7 @@
     document.getElementById('m-chunks').textContent = fmt(chunks);
     document.getElementById('m-memories').textContent = fmt(memories);
     document.getElementById('m-projects').textContent = fmt(d.project_count || 0);
+    document.getElementById('m-global').textContent = fmt(d.global_memories || 0);
     document.getElementById('m-recalls').textContent = fmt(d.total_recalls);
     if (d.total_recalls > 0) {
       document.getElementById('m-recalls-sub').textContent =
@@ -695,6 +738,39 @@
     });
   }
 
+  /* ── Quality ───────────────────────────────────────────── */
+  function renderQuality(data) {
+    var section = document.getElementById('quality-section');
+    var tbody = document.getElementById('quality-body');
+    tbody.innerHTML = '';
+    if (!data || !data.length) { section.style.display = 'none'; return; }
+    section.style.display = '';
+    data.forEach(function(row) {
+      var neg = row.negative_count || 0;
+      var score = row.quality_score || 0;
+      var barClass = score >= 0.5 ? 'quality-fill-ok' : (score >= 0.25 ? 'quality-fill-warn' : 'quality-fill-error');
+      var barWidth = Math.max(2, Math.round(score * 100));
+      var projLabel = row.project === '_global'
+        ? '<span class="badge badge-global">Global</span>'
+        : esc(row.project);
+      var typeLabel = esc(row.mem_type || row.kind);
+      var negHtml = neg > 0
+        ? '<span style="color:var(--error);font-weight:600;">' + neg + '</span>'
+        : '<span style="color:var(--text-muted);">0</span>';
+      var tr = el('tr', {});
+      tr.innerHTML =
+        '<td class="truncate" style="max-width:260px" title="' + esc(row.preview) + '">' + esc(row.preview) + '</td>' +
+        '<td>' + projLabel + '</td>' +
+        '<td><span class="mono" style="font-size:11px;color:var(--text-muted);">' + typeLabel + '</span></td>' +
+        '<td>' + (row.recall_count || 0) + '</td>' +
+        '<td style="color:var(--ok);">' + (row.use_count || 0) + '</td>' +
+        '<td>' + negHtml + '</td>' +
+        '<td class="mono">' + fmtScore(score) + '</td>' +
+        '<td><div class="quality-bar"><div class="quality-fill ' + barClass + '" style="width:' + barWidth + '%;"></div></div></td>';
+      tbody.appendChild(tr);
+    });
+  }
+
   /* ── Projects ──────────────────────────────────────────── */
   function renderProjects(data) {
     projectsData = data;
@@ -737,9 +813,13 @@
     sorted.forEach(function(row) {
       var tr = el('tr', {});
       if (projectFilter === row.project) tr.classList.add('selected');
+      var isGlobal = row.project === '_global';
+      var projCell = isGlobal
+        ? '<td><span class="badge badge-global">Global</span></td>'
+        : '<td class="truncate mono" title="' + esc(row.project) + '">' + esc(row.project) + '</td>';
       if (isRecallMode) {
         tr.innerHTML =
-          '<td class="truncate mono" title="' + esc(row.project) + '">' + esc(row.project) + '</td>' +
+          projCell +
           '<td>' + fmt(row.recalls) + '</td><td>' + fmt(row.tokens) + '</td>' +
           '<td>' + fmtScore(row.avg_score) + '</td>' +
           '<td style="color:var(--ok)">' + (row.ok_count||0) + '</td>' +
@@ -747,10 +827,10 @@
           '<td style="color:var(--warn)">' + (row.extractive_count||0) + '</td>';
       } else {
         var hasMemories = (row.memories||0) > 0;
-        var badgeCls = hasMemories ? 'badge-ok' : 'badge-extractive';
-        var status = hasMemories ? 'Active' : 'Ingested';
+        var badgeCls = isGlobal ? 'badge-global' : (hasMemories ? 'badge-ok' : 'badge-extractive');
+        var status = isGlobal ? 'Global' : (hasMemories ? 'Active' : 'Ingested');
         tr.innerHTML =
-          '<td class="truncate mono" title="' + esc(row.project) + '">' + esc(row.project) + '</td>' +
+          projCell +
           '<td>' + fmt(row.chunks||0) + '</td><td>' + fmt(row.memories||0) + '</td>' +
           '<td>' + fmt(row.total_tokens||0) + '</td>' +
           '<td><span class="badge ' + badgeCls + '">' + status + '</span></td>' +
@@ -931,6 +1011,7 @@
   async function loadTrend()      { try { renderTrend(await api('/api/recall-trend?days=30')); } catch(e) { console.warn('trend',e); } }
   async function loadROI()        { try { renderROI(await api('/api/roi')); } catch(e) { console.warn('roi',e); } }
   async function loadROITrend()   { try { renderROITrend(await api('/api/roi-trend')); } catch(e) { console.warn('roi-trend',e); } }
+  async function loadQuality()    { try { renderQuality(await api('/api/quality')); } catch(e) { console.warn('quality',e); } }
   async function loadRecalls() {
     try {
       var url = '/api/recalls?limit=50' + (projectFilter ? '&project=' + encodeURIComponent(projectFilter) : '');
@@ -939,7 +1020,7 @@
   }
 
   async function refresh() {
-    await Promise.allSettled([loadSummary(), loadProjects(), loadEfficiency(), loadSessionEfficiency(), loadHealth(), loadTrend(), loadRecalls(), loadROI(), loadROITrend()]);
+    await Promise.allSettled([loadSummary(), loadProjects(), loadEfficiency(), loadSessionEfficiency(), loadHealth(), loadTrend(), loadRecalls(), loadROI(), loadROITrend(), loadQuality()]);
     document.getElementById('last-updated').textContent = new Date().toLocaleTimeString();
 
     renderMiniBars('mb-chunks', null);
@@ -956,6 +1037,12 @@
     });
   });
 
+  (async function() {
+    try {
+      var v = await api('/api/version');
+      document.getElementById('version-badge').textContent = 'v' + v.version;
+    } catch(e) {}
+  })();
   refresh();
   setInterval(refresh, 30000);
 })();

{memor_cli-0.5.0 → memor_cli-0.6.0}/memor/feedback.py

@@ -19,9 +19,29 @@ _MIN_WORDS = 4
 _MATCH_RATIO = 0.10
 _SEMANTIC_SIM_THRESHOLD = 0.45
 
-
-def _extract_assistant_texts(transcript_path: Path) -> list[str]:
-    texts = []
+_REJECTION_PATTERNS = [
+    "no that's wrong", "that's not right", "that's incorrect", "that's outdated",
+    "no, we", "no we", "actually we", "actually, we",
+    "we switched", "we moved", "we changed", "we no longer",
+    "that's not how", "not what i meant", "wrong approach",
+    "we don't use", "we stopped using", "we dropped",
+]
+
+_CONTRADICTION_PATTERNS = [
+    "however, looking at the current code",
+    "but the current implementation",
+    "actually use", "actually uses",
+    "instead of what the memory",
+    "contrary to the recalled",
+    "the memory is outdated",
+    "this is no longer",
+]
+
+
+def _extract_texts(transcript_path: Path) -> tuple[list[str], list[str]]:
+    """Extract (assistant_texts, user_texts) from a transcript JSONL."""
+    assistant_texts = []
+    user_texts = []
     for line in transcript_path.read_text().splitlines():
         line = line.strip()
         if not line:
@@ -30,17 +50,40 @@ def _extract_assistant_texts(transcript_path: Path) -> list[str]:
             rec = json.loads(line)
         except json.JSONDecodeError:
             continue
-        if rec.get("type") != "assistant":
-            continue
         msg = rec.get("message", {})
         content = msg.get("content", "")
+        parts = []
         if isinstance(content, str):
-            texts.append(content.lower())
+            parts.append(content.lower())
         elif isinstance(content, list):
             for block in content:
                 if isinstance(block, dict) and block.get("type") == "text":
-                    texts.append(block.get("text", "").lower())
-    return texts
+                    parts.append(block.get("text", "").lower())
+        if rec.get("type") == "assistant":
+            assistant_texts.extend(parts)
+        elif rec.get("type") == "human":
+            user_texts.extend(parts)
+    return assistant_texts, user_texts
+
+
+def _extract_assistant_texts(transcript_path: Path) -> list[str]:
+    assistant_texts, _ = _extract_texts(transcript_path)
+    return assistant_texts
+
+
+def _detect_negative_signals(assistant_texts: list[str], user_texts: list[str]) -> bool:
+    """Detect if user rejected or assistant contradicted recalled content."""
+    combined_user = " ".join(user_texts)
+    for pattern in _REJECTION_PATTERNS:
+        if pattern in combined_user:
+            return True
+
+    combined_assistant = " ".join(assistant_texts)
+    for pattern in _CONTRADICTION_PATTERNS:
+        if pattern in combined_assistant:
+            return True
+
+    return False
 
 
 def _text_was_used(memory_text: str, assistant_texts: list[str]) -> bool:
@@ -106,7 +149,7 @@ def analyze_session_feedback(
     if not recalled_ids:
         return 0
 
-    assistant_texts = _extract_assistant_texts(transcript_path)
+    assistant_texts, user_texts = _extract_texts(transcript_path)
     if not assistant_texts:
         return 0
 
@@ -126,4 +169,7 @@ def analyze_session_feedback(
     if used_ids:
         store.record_usage(used_ids)
 
+    if _detect_negative_signals(assistant_texts, user_texts) and recalled_ids:
+        store.record_negative(list(recalled_ids))
+
     return len(used_ids)

memor_cli-0.6.0/memor/global_memories.py

@@ -0,0 +1,102 @@
+"""Cross-project global memory promotion.
+
+Detects patterns that appear across 3+ projects and promotes them to the
+_global scope so they're recalled everywhere. Fully local — no API keys.
+"""
+from __future__ import annotations
+import hashlib
+import math
+from memor.store.sqlite_store import SqliteStore
+from memor.types import Artifact, Scope, GLOBAL_PROJECT
+from memor.tokencount import count_tokens
+
+
+def _cosine(a: list[float], b: list[float]) -> float:
+    dot = sum(x * y for x, y in zip(a, b))
+    na = math.sqrt(sum(x * x for x in a))
+    nb = math.sqrt(sum(x * x for x in b))
+    return dot / (na * nb) if na and nb else 0.0
+
+
+def find_promotion_candidates(
+    store: SqliteStore, embedder, *,
+    min_projects: int = 3, sim_threshold: float = 0.85,
+) -> list[dict]:
+    """Find memories that appear in min_projects+ different projects.
+
+    Returns list of dicts: {"text": str, "source_ids": [str], "projects": [str]}
+    """
+    rows = store.db.execute(
+        "SELECT * FROM artifacts WHERE kind='memory' AND active=1 AND project != ?",
+        (GLOBAL_PROJECT,)
+    ).fetchall()
+    if len(rows) < min_projects:
+        return []
+
+    memories = [store._row_to_artifact(r) for r in rows]
+    vecs = embedder.embed([m.text for m in memories])
+
+    clusters: list[dict] = []
+    assigned = set()
+
+    for i, mem in enumerate(memories):
+        if i in assigned:
+            continue
+        cluster = {"text": mem.text, "source_ids": [mem.id], "projects": {mem.project}}
+        assigned.add(i)
+        for j in range(i + 1, len(memories)):
+            if j in assigned:
+                continue
+            if memories[j].project in cluster["projects"]:
+                continue
+            sim = _cosine(vecs[i], vecs[j])
+            if sim >= sim_threshold:
+                cluster["source_ids"].append(memories[j].id)
+                cluster["projects"].add(memories[j].project)
+                assigned.add(j)
+        if len(cluster["projects"]) >= min_projects:
+            cluster["projects"] = sorted(cluster["projects"])
+            clusters.append(cluster)
+
+    return clusters
+
+
+def promote_to_global(
+    store: SqliteStore, embedder,
+    text: str, source_ids: list[str],
+) -> str | None:
+    """Create a _global memory from text and deactivate source duplicates.
+
+    Returns the new global memory ID, or None if a similar global already exists.
+    """
+    vec = embedder.embed([text])[0]
+
+    existing = store.search(vec, Scope(project=GLOBAL_PROJECT, kinds=["memory"]), k=1)
+    if existing and existing[0][1] >= 0.85:
+        return None
+
+    mid = f"mem:global:{hashlib.sha1(text.encode()).hexdigest()[:8]}"
+    import time
+    art = Artifact(
+        id=mid, kind="memory", project=GLOBAL_PROJECT, source="promotion",
+        text=text, token_count=max(1, count_tokens(text)), created_at=time.time(),
+        meta={"mem_type": "global", "promoted_from": source_ids},
+    )
+    store.add_artifacts([art], [vec])
+
+    for sid in source_ids:
+        store.deactivate(sid, superseded_by=mid)
+
+    return mid
+
+
+def run_promotion(store: SqliteStore, embedder, *, min_projects: int = 3) -> int:
+    """Find and promote cross-project patterns. Returns count promoted."""
+    candidates = find_promotion_candidates(
+        store, embedder, min_projects=min_projects)
+    promoted = 0
+    for c in candidates:
+        mid = promote_to_global(store, embedder, c["text"], c["source_ids"])
+        if mid:
+            promoted += 1
+    return promoted