memor-cli 0.8.0__tar.gz → 0.9.0__tar.gz

{memor_cli-0.8.0/memor_cli.egg-info → memor_cli-0.9.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memor-cli
-Version: 0.8.0
+Version: 0.9.0
 Summary: Measured memory for coding agents. Fire and forget — no API keys needed.
 Author-email: Nimit Bhandari <nimitbhandari17@gmail.com>
 License-Expression: MIT
@@ -50,9 +50,9 @@ Dynamic: license-file
 [![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/)
 
-**Automatic background memory for Claude Code, Codex, and Copilot.** Fire and forget — no API keys needed.
+**Automatic background memory for Claude Code, Cursor, Codex, and Copilot.** Fire and forget — no API keys needed.
 
-Memor watches your coding sessions, extracts decisions and patterns, and recalls relevant context on every prompt. Works with Claude Code, OpenAI Codex CLI, and GitHub Copilot CLI. Zero configuration. One install. Your agent remembers everything.
+Memor watches your coding sessions, extracts decisions and patterns, and recalls relevant context on every prompt. Works with Claude Code, Cursor, OpenAI Codex CLI, and GitHub Copilot CLI. Zero configuration. One install. Your agent remembers everything.
 
 ---
 
@@ -119,15 +119,16 @@ memor dashboard
 | **Claude Code** | `UserPromptSubmit` + `additionalContext` | `~/.claude/settings.json` | `memor install-hook --agent claude` |
 | **Codex CLI** | `UserPromptSubmit` + `additionalContext` | `~/.codex/hooks/hooks.json` | `memor install-hook --agent codex` |
 | **Copilot CLI** | `userPromptSubmitted` + `additionalContext` | `~/.copilot/hooks/memor.json` | `memor install-hook --agent copilot` |
+| **Cursor** | `beforeSubmitPrompt` + `additionalContext` | `~/.claude/settings.json` (loaded as Claude user hooks) | automatic — covered by the Claude install |
 
-A single `memor-hook` binary auto-detects which agent is calling it — no separate entry points needed. The dashboard tracks recalls per agent so you can see usage across all your environments.
+A single `memor-hook` binary auto-detects which agent is calling it — no separate entry points needed. Cursor loads the same Claude user hooks, so installing for Claude Code covers Cursor too. The dashboard tracks recalls per agent so you can see usage across all your environments.
 
 > **Note:** Cloud-hosted agents (Codex cloud, Copilot cloud agent) run in remote sandboxes and cannot reach local hooks. MCP server support for sandboxed agents is planned ([#26](https://github.com/bnimit/memor-ai/issues/26)).
 
 **Two background processes:**
 
 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, auto-compacts the vector index when bloated, tracks session-level token usage. All local.
-2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms. Works across Claude Code, Codex, and Copilot.
+2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms. Works across Claude Code, Cursor, Codex, and Copilot.
 
 **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.
 
@@ -201,7 +202,7 @@ memor dashboard
 
 Dark fintech-inspired UI showing:
 - **Hero metrics** — total memories, recall count, avg latency, coverage — with sparkline bars
-- **Agent breakdown** — per-agent recall stats (Claude, Codex, Copilot) with hit rates
+- **Agent breakdown** — per-agent recall stats (Claude, Cursor, Codex, Copilot) with hit rates
 - **Daily recall activity** — stacked bar chart of hits vs misses over time
 - **Session efficiency** — real token savings measured from API usage data (avg tokens/turn with vs without recall)
 - **Per-project breakdown** — artifact counts, token totals, last activity
@@ -283,7 +284,7 @@ memor/
     +-- judge.py           LLM-as-judge evaluation
     +-- embed_benchmark.py Embedding model comparison
 
-memor/hook_cli.py          Hook entry point — auto-detects Claude/Codex/Copilot
+memor/hook_cli.py          Hook entry point — auto-detects Claude/Cursor/Codex/Copilot
 memor/hook_server.py       Hook server with agent detection + response formatting
 skill/recall.py            Standalone recall script
 ```
@@ -331,7 +332,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 298 tests
+pytest  # 304 tests
 ```
 
 ---

{memor_cli-0.8.0 → memor_cli-0.9.0}/README.md

@@ -13,9 +13,9 @@
 [![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/)
 
-**Automatic background memory for Claude Code, Codex, and Copilot.** Fire and forget — no API keys needed.
+**Automatic background memory for Claude Code, Cursor, Codex, and Copilot.** Fire and forget — no API keys needed.
 
-Memor watches your coding sessions, extracts decisions and patterns, and recalls relevant context on every prompt. Works with Claude Code, OpenAI Codex CLI, and GitHub Copilot CLI. Zero configuration. One install. Your agent remembers everything.
+Memor watches your coding sessions, extracts decisions and patterns, and recalls relevant context on every prompt. Works with Claude Code, Cursor, OpenAI Codex CLI, and GitHub Copilot CLI. Zero configuration. One install. Your agent remembers everything.
 
 ---
 
@@ -82,15 +82,16 @@ memor dashboard
 | **Claude Code** | `UserPromptSubmit` + `additionalContext` | `~/.claude/settings.json` | `memor install-hook --agent claude` |
 | **Codex CLI** | `UserPromptSubmit` + `additionalContext` | `~/.codex/hooks/hooks.json` | `memor install-hook --agent codex` |
 | **Copilot CLI** | `userPromptSubmitted` + `additionalContext` | `~/.copilot/hooks/memor.json` | `memor install-hook --agent copilot` |
+| **Cursor** | `beforeSubmitPrompt` + `additionalContext` | `~/.claude/settings.json` (loaded as Claude user hooks) | automatic — covered by the Claude install |
 
-A single `memor-hook` binary auto-detects which agent is calling it — no separate entry points needed. The dashboard tracks recalls per agent so you can see usage across all your environments.
+A single `memor-hook` binary auto-detects which agent is calling it — no separate entry points needed. Cursor loads the same Claude user hooks, so installing for Claude Code covers Cursor too. The dashboard tracks recalls per agent so you can see usage across all your environments.
 
 > **Note:** Cloud-hosted agents (Codex cloud, Copilot cloud agent) run in remote sandboxes and cannot reach local hooks. MCP server support for sandboxed agents is planned ([#26](https://github.com/bnimit/memor-ai/issues/26)).
 
 **Two background processes:**
 
 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, auto-compacts the vector index when bloated, tracks session-level token usage. All local.
-2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms. Works across Claude Code, Codex, and Copilot.
+2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms. Works across Claude Code, Cursor, Codex, and Copilot.
 
 **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.
 
@@ -164,7 +165,7 @@ memor dashboard
 
 Dark fintech-inspired UI showing:
 - **Hero metrics** — total memories, recall count, avg latency, coverage — with sparkline bars
-- **Agent breakdown** — per-agent recall stats (Claude, Codex, Copilot) with hit rates
+- **Agent breakdown** — per-agent recall stats (Claude, Cursor, Codex, Copilot) with hit rates
 - **Daily recall activity** — stacked bar chart of hits vs misses over time
 - **Session efficiency** — real token savings measured from API usage data (avg tokens/turn with vs without recall)
 - **Per-project breakdown** — artifact counts, token totals, last activity
@@ -246,7 +247,7 @@ memor/
     +-- judge.py           LLM-as-judge evaluation
     +-- embed_benchmark.py Embedding model comparison
 
-memor/hook_cli.py          Hook entry point — auto-detects Claude/Codex/Copilot
+memor/hook_cli.py          Hook entry point — auto-detects Claude/Cursor/Codex/Copilot
 memor/hook_server.py       Hook server with agent detection + response formatting
 skill/recall.py            Standalone recall script
 ```
@@ -294,7 +295,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 298 tests
+pytest  # 304 tests
 ```
 
 ---

memor_cli-0.9.0/memor/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.9.0"

{memor_cli-0.8.0 → memor_cli-0.9.0}/memor/dashboard/static/index.html

@@ -310,6 +310,8 @@
   .badge-codex::before { background: #3dd68c; }
   .badge-copilot { background: rgba(100,160,255,0.12); color: #64a0ff; }
   .badge-copilot::before { background: #64a0ff; }
+  .badge-cursor { background: rgba(192,132,252,0.12); color: #c084fc; }
+  .badge-cursor::before { background: #c084fc; }
 
   .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; }
@@ -613,6 +615,7 @@
       claude:  ['badge-claude',  'Claude'],
       codex:   ['badge-codex',   'Codex'],
       copilot: ['badge-copilot', 'Copilot'],
+      cursor:  ['badge-cursor',  'Cursor'],
     };
     var m = map[agent || 'claude'] || ['badge-claude', agent || 'claude'];
     return '<span class="badge ' + m[0] + '">' + m[1] + '</span>';
@@ -1033,7 +1036,7 @@
     if (!data || data.length <= 1) { section.style.display = 'none'; return; }
     section.style.display = '';
     grid.innerHTML = '';
-    var colors = { claude: 'var(--accent)', codex: '#3dd68c', copilot: '#64a0ff' };
+    var colors = { claude: 'var(--accent)', codex: '#3dd68c', copilot: '#64a0ff', cursor: '#c084fc' };
     data.forEach(function(row) {
       var agent = row.agent || 'claude';
       var color = colors[agent] || 'var(--text-muted)';

{memor_cli-0.8.0 → memor_cli-0.9.0}/memor/hook_server.py

@@ -17,11 +17,17 @@ def detect_agent(req: dict) -> str:
     event = req.get("hook_event_name", "")
     if event == "userPromptSubmitted":
         return "copilot"
-    # A real Codex payload carries both `model` and a Codex-specific `turn_id`
-    # extension; Claude Code sends neither. We match on EITHER (not both) so
-    # detection still holds if a Codex version drops one field. Do NOT key on
-    # `permission_mode`: it is a base field on every Claude Code hook input, so
-    # it cannot discriminate between the two.
+    # Cursor fires a `beforeSubmitPrompt` hook and stamps every payload with a
+    # `cursor_version` field. It must be matched BEFORE the codex check below,
+    # because Cursor also sends `model` (e.g. "composer-2.5-fast") which would
+    # otherwise be mistaken for Codex.
+    if event == "beforeSubmitPrompt" or "cursor_version" in req:
+        return "cursor"
+    # A real Codex payload carries `model` and/or a Codex-specific `turn_id`
+    # extension; Claude Code sends neither. We match on EITHER so detection still
+    # holds if a Codex version drops one field. Do NOT key on `permission_mode`:
+    # it is a base field on every Claude Code hook input, so it cannot
+    # discriminate between the two.
     if "turn_id" in req or "model" in req:
         return "codex"
     return "claude"
@@ -68,6 +74,12 @@ def handle_request(req: dict, *, db_path: str = DEFAULT_DB,
     from memor.project import resolve_project
 
     cwd = req.get("cwd", "")
+    if not cwd:
+        # Cursor sends `workspace_roots` (a list) instead of `cwd`; fall back to
+        # the first root so the recall is scoped to the real project.
+        roots = req.get("workspace_roots") or []
+        if roots:
+            cwd = roots[0]
     project = resolve_project(cwd) if cwd else "unknown"
     query = req.get("prompt", "")
     session_id = req.get("session_id", "")

memor_cli-0.9.0/memor/llm/openai_compat.py

@@ -0,0 +1,27 @@
+import httpx
+
+
+class OpenAICompatLLM:
+    def __init__(self, base_url: str, api_key: str, model: str,
+                 temperature: float | None = None):
+        self.base_url, self.api_key, self.model = base_url, api_key, model
+        # temperature=0 makes the eval judge deterministic/repeatable; None omits
+        # the field (server default) for normal use.
+        self.temperature = temperature
+
+    def complete(self, prompt: str, *, max_tokens: int = 1024) -> str:
+        payload = {
+            "model": self.model,
+            "max_tokens": max_tokens,
+            "messages": [{"role": "user", "content": prompt}],
+        }
+        if self.temperature is not None:
+            payload["temperature"] = self.temperature
+        r = httpx.post(
+            f"{self.base_url}/chat/completions",
+            headers={"Authorization": f"Bearer {self.api_key}"},
+            json=payload,
+            timeout=60,
+        )
+        r.raise_for_status()
+        return r.json()["choices"][0]["message"]["content"]

{memor_cli-0.8.0 → memor_cli-0.9.0}/memor/retrieve/retriever.py

@@ -86,8 +86,10 @@ class Retriever:
         rel_min = min(rel_vals) if rel_vals else 0.0
         rel_range = ((max(rel_vals) - rel_min) if rel_vals else 1.0) or 1.0
 
-        quality_cache = {}
-        has_quality = hasattr(self.store, 'get_quality_score')
+        if hasattr(self.store, 'get_quality_scores'):
+            quality_scores = self.store.get_quality_scores(list(arts_by_id))
+        else:
+            quality_scores = {}
 
         for aid, a in arts_by_id.items():
             norm_rel = (fused.get(aid, 0.0) - rel_min) / rel_range
@@ -97,9 +99,7 @@ class Retriever:
 
             kind_boost = KIND_WEIGHTS.get(a.kind, 1.0) - 1.0
 
-            if has_quality and aid not in quality_cache:
-                quality_cache[aid] = self.store.get_quality_score(aid)
-            quality = quality_cache.get(aid, 0.5)
+            quality = quality_scores.get(aid, 0.5)
 
             score = (self.w_sim * norm_rel + self.w_rec * recency
                      + self.w_kind * kind_boost + self.w_qual * quality)

{memor_cli-0.8.0 → memor_cli-0.9.0}/memor/store/sqlite_store.py

@@ -170,8 +170,13 @@ class SqliteStore:
                         text=r["text"], token_count=r["token_count"], created_at=r["created_at"],
                         meta=json.loads(r["meta"]))
 
+    # sqlite-vec caps the KNN `k` parameter at this value.
+    _VEC_KNN_LIMIT = 4096
+
     def search(self, vector: list[float], scope: Scope, k: int) -> list[tuple[Artifact, float]]:
         from memor.types import GLOBAL_PROJECT
+        # Cap the KNN fetch at sqlite-vec's hard limit so large k can't error.
+        fetch = min(max(k * 20, 200), self._VEC_KNN_LIMIT)
         rows = self.db.execute(f"""
           SELECT a.*, v.distance AS distance
           FROM (SELECT rowid, distance FROM vec_artifacts
@@ -182,7 +187,7 @@ class SqliteStore:
             AND (? IS NULL OR a.created_at >= ?)
             AND (? IS NULL OR a.created_at <= ?)
           ORDER BY v.distance ASC
-        """, (_serialize(vector), max(k*20, 200),
+        """, (_serialize(vector), fetch,
               scope.project, scope.project, GLOBAL_PROJECT,
               scope.since, scope.since,
               scope.until, scope.until)).fetchall()
@@ -457,6 +462,19 @@ class SqliteStore:
             (artifact_id,)).fetchone()
         return row["quality_score"] if row else 0.5
 
+    def get_quality_scores(self, artifact_ids: list[str]) -> dict[str, float]:
+        """Batch form of get_quality_score: one query for the whole candidate
+        set (callers default missing ids to 0.5). Keeps the widened retrieval
+        pool off the N+1-query path on the per-prompt hot path."""
+        ids = list(artifact_ids)
+        if not ids:
+            return {}
+        qmarks = ",".join("?" * len(ids))
+        rows = self.db.execute(
+            f"SELECT artifact_id, quality_score FROM memory_quality "
+            f"WHERE artifact_id IN ({qmarks})", ids).fetchall()
+        return {r["artifact_id"]: r["quality_score"] for r in rows}
+
     def get_stale_memories(self, days: int = 30) -> list[str]:
         import time as _time
         cutoff = _time.time() - (days * 86400)

{memor_cli-0.8.0 → memor_cli-0.9.0/memor_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memor-cli
-Version: 0.8.0
+Version: 0.9.0
 Summary: Measured memory for coding agents. Fire and forget — no API keys needed.
 Author-email: Nimit Bhandari <nimitbhandari17@gmail.com>
 License-Expression: MIT
@@ -50,9 +50,9 @@ Dynamic: license-file
 [![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/)
 
-**Automatic background memory for Claude Code, Codex, and Copilot.** Fire and forget — no API keys needed.
+**Automatic background memory for Claude Code, Cursor, Codex, and Copilot.** Fire and forget — no API keys needed.
 
-Memor watches your coding sessions, extracts decisions and patterns, and recalls relevant context on every prompt. Works with Claude Code, OpenAI Codex CLI, and GitHub Copilot CLI. Zero configuration. One install. Your agent remembers everything.
+Memor watches your coding sessions, extracts decisions and patterns, and recalls relevant context on every prompt. Works with Claude Code, Cursor, OpenAI Codex CLI, and GitHub Copilot CLI. Zero configuration. One install. Your agent remembers everything.
 
 ---
 
@@ -119,15 +119,16 @@ memor dashboard
 | **Claude Code** | `UserPromptSubmit` + `additionalContext` | `~/.claude/settings.json` | `memor install-hook --agent claude` |
 | **Codex CLI** | `UserPromptSubmit` + `additionalContext` | `~/.codex/hooks/hooks.json` | `memor install-hook --agent codex` |
 | **Copilot CLI** | `userPromptSubmitted` + `additionalContext` | `~/.copilot/hooks/memor.json` | `memor install-hook --agent copilot` |
+| **Cursor** | `beforeSubmitPrompt` + `additionalContext` | `~/.claude/settings.json` (loaded as Claude user hooks) | automatic — covered by the Claude install |
 
-A single `memor-hook` binary auto-detects which agent is calling it — no separate entry points needed. The dashboard tracks recalls per agent so you can see usage across all your environments.
+A single `memor-hook` binary auto-detects which agent is calling it — no separate entry points needed. Cursor loads the same Claude user hooks, so installing for Claude Code covers Cursor too. The dashboard tracks recalls per agent so you can see usage across all your environments.
 
 > **Note:** Cloud-hosted agents (Codex cloud, Copilot cloud agent) run in remote sandboxes and cannot reach local hooks. MCP server support for sandboxed agents is planned ([#26](https://github.com/bnimit/memor-ai/issues/26)).
 
 **Two background processes:**
 
 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, auto-compacts the vector index when bloated, tracks session-level token usage. All local.
-2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms. Works across Claude Code, Codex, and Copilot.
+2. **Hook** — fires on every prompt, recalls relevant memories, injects them as context. Sub-15ms. Works across Claude Code, Cursor, Codex, and Copilot.
 
 **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.
 
@@ -201,7 +202,7 @@ memor dashboard
 
 Dark fintech-inspired UI showing:
 - **Hero metrics** — total memories, recall count, avg latency, coverage — with sparkline bars
-- **Agent breakdown** — per-agent recall stats (Claude, Codex, Copilot) with hit rates
+- **Agent breakdown** — per-agent recall stats (Claude, Cursor, Codex, Copilot) with hit rates
 - **Daily recall activity** — stacked bar chart of hits vs misses over time
 - **Session efficiency** — real token savings measured from API usage data (avg tokens/turn with vs without recall)
 - **Per-project breakdown** — artifact counts, token totals, last activity
@@ -283,7 +284,7 @@ memor/
     +-- judge.py           LLM-as-judge evaluation
     +-- embed_benchmark.py Embedding model comparison
 
-memor/hook_cli.py          Hook entry point — auto-detects Claude/Codex/Copilot
+memor/hook_cli.py          Hook entry point — auto-detects Claude/Cursor/Codex/Copilot
 memor/hook_server.py       Hook server with agent detection + response formatting
 skill/recall.py            Standalone recall script
 ```
@@ -331,7 +332,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 298 tests
+pytest  # 304 tests
 ```
 
 ---

{memor_cli-0.8.0 → memor_cli-0.9.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "memor-cli"
-version = "0.8.0"
+version = "0.9.0"
 description = "Measured memory for coding agents. Fire and forget — no API keys needed."
 readme = "README.md"
 license = "MIT"

{memor_cli-0.8.0 → memor_cli-0.9.0}/tests/test_multi_agent_hook.py

@@ -44,6 +44,34 @@ def test_detect_copilot():
     assert detect_agent(req) == "copilot"
 
 
+def test_detect_cursor():
+    # Real Cursor 3.7.36 beforeSubmitPrompt payload (captured from cursor.hooks
+    # log). Note it carries `model` ("composer-2.5-fast") and a unique
+    # `cursor_version`, and uses `workspace_roots` instead of `cwd`.
+    req = {"conversation_id": "c1", "generation_id": "g1",
+           "model": "composer-2.5-fast", "composer_mode": "agent",
+           "prompt": "kill the background terminals please", "session_id": "c1",
+           "hook_event_name": "beforeSubmitPrompt", "cursor_version": "3.7.36",
+           "workspace_roots": ["/proj"]}
+    assert detect_agent(req) == "cursor"
+
+
+def test_detect_cursor_by_version_field():
+    # Defensive: cursor_version alone identifies Cursor even if the event name
+    # ever changes.
+    req = {"prompt": "x", "session_id": "c1", "cursor_version": "3.8.0",
+           "hook_event_name": "UserPromptSubmit", "model": "composer-2.5"}
+    assert detect_agent(req) == "cursor"
+
+
+def test_cursor_model_is_not_mistaken_for_codex():
+    # Regression: Cursor sends `model`, but it must NOT fall into the codex
+    # branch — the cursor markers win.
+    req = {"prompt": "x", "hook_event_name": "beforeSubmitPrompt",
+           "model": "composer-2.5-fast", "workspace_roots": ["/proj"]}
+    assert detect_agent(req) == "cursor"
+
+
 def test_detect_unknown_defaults_to_claude():
     req = {"cwd": "/proj", "prompt": "hello"}
     assert detect_agent(req) == "claude"
@@ -115,6 +143,22 @@ def test_handle_request_copilot_format(tmp_path):
     assert "hookSpecificOutput" not in resp
 
 
+def test_handle_request_cursor_resolves_project_from_workspace_roots(tmp_path):
+    """Cursor sends `workspace_roots` instead of `cwd`. handle_request must fall
+    back to it so the recall is scoped to the real project (not 'unknown') and
+    logged as the cursor agent."""
+    db_path, e = _make_db(tmp_path)
+    req = {"prompt": "password hashing", "session_id": "c1",
+           "hook_event_name": "beforeSubmitPrompt", "cursor_version": "3.7.36",
+           "model": "composer-2.5-fast", "workspace_roots": ["/proj"]}
+    handle_request(req, db_path=db_path, embedder=e)
+    store = SqliteStore(db_path, dim=16)
+    row = store.db.execute(
+        "SELECT project, agent FROM recall_log ORDER BY id DESC LIMIT 1").fetchone()
+    assert row["project"] == "proj"
+    assert row["agent"] == "cursor"
+
+
 # --- install-hook config generation ---
 
 def test_install_hook_codex(tmp_path):

{memor_cli-0.8.0 → memor_cli-0.9.0}/tests/test_store.py

@@ -32,3 +32,25 @@ def test_deactivate_excludes_from_search(tmp_path):
     s.deactivate("old", superseded_by="new")
     ids = [a.id for a, _ in s.search(e.embed(["use library"])[0], Scope(project="p"), k=5)]
     assert "old" not in ids and "new" in ids
+
+
+# --- KNN-fetch cap + batched quality lookup (safe retrieval wins) ---
+
+def test_search_knn_fetch_capped(tmp_path):
+    # A large k must not exceed sqlite-vec's internal knn limit (4096).
+    e = FakeEmbedder(dim=16)
+    s = SqliteStore(str(tmp_path / "m.db"), dim=16)
+    s.add_artifacts([make("a", "p", "hello world", 1)], e.embed(["hello world"]))
+    hits = s.search(e.embed(["hello world"])[0], Scope(project="p"), k=300)
+    assert len(hits) <= 1  # no OperationalError, returns what's available
+
+
+def test_get_quality_scores_batch_matches_per_id(tmp_path):
+    s = SqliteStore(str(tmp_path / "m.db"), dim=16)
+    s.record_recall(["a", "b"])
+    s.record_usage(["a"])
+    scores = s.get_quality_scores(["a", "b", "missing"])
+    assert scores["a"] == s.get_quality_score("a")
+    assert scores["b"] == s.get_quality_score("b")
+    assert "missing" not in scores
+    assert s.get_quality_scores([]) == {}

memor_cli-0.8.0/memor/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.8.0"

memor_cli-0.8.0/memor/llm/openai_compat.py

@@ -1,20 +0,0 @@
-import httpx
-
-
-class OpenAICompatLLM:
-    def __init__(self, base_url: str, api_key: str, model: str):
-        self.base_url, self.api_key, self.model = base_url, api_key, model
-
-    def complete(self, prompt: str, *, max_tokens: int = 1024) -> str:
-        r = httpx.post(
-            f"{self.base_url}/chat/completions",
-            headers={"Authorization": f"Bearer {self.api_key}"},
-            json={
-                "model": self.model,
-                "max_tokens": max_tokens,
-                "messages": [{"role": "user", "content": prompt}],
-            },
-            timeout=60,
-        )
-        r.raise_for_status()
-        return r.json()["choices"][0]["message"]["content"]