memor-cli 0.7.1__tar.gz → 0.9.0__tar.gz

{memor_cli-0.7.1/memor_cli.egg-info → memor_cli-0.9.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memor-cli
-Version: 0.7.1
+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.
 
 ---
 
@@ -73,7 +73,7 @@ memor service install
 memor daemon
 ```
 
-That's it. Every prompt now gets automatic context recall. Open the dashboard to see it working:
+That's it. Every prompt now gets automatic context recall. `memor service install` also starts the dashboard as a background service, so it's already live at http://localhost:8420 (and is recycled whenever you stop/restart/uninstall the service). To run it in the foreground instead:
 
 ```bash
 memor dashboard
@@ -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
@@ -218,10 +219,13 @@ memor install-hook                   Install hook + download model (interactive
 memor daemon                         Auto-ingest + distill (background watcher)
 memor dashboard                      Web dashboard on localhost:8420
 memor version                        Print installed version
-memor service install                Run daemon as background service (launchd/systemd)
-memor service stop                   Stop the background service
-memor service uninstall              Remove the background service
-memor service status                 Check if the service is running
+memor service install                Run daemon + dashboard as background services (launchd/systemd)
+  --no-dashboard                     Install only the daemon
+memor service restart                Restart both services (use after `pipx upgrade`)
+memor service stop                   Stop both background services
+memor service uninstall              Remove both background services
+memor service status                 Show daemon + dashboard status
+  (dashboard port: set MEMOR_DASHBOARD_PORT, default 8420)
 memor query <text>                   Search memories from the CLI
 memor reingest                       Wipe DB and re-ingest everything
 memor reingest --project <name>      Re-ingest only one project
@@ -280,7 +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
 ```
@@ -328,7 +332,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 287 tests
+pytest  # 304 tests
 ```
 
 ---

{memor_cli-0.7.1 → 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.
 
 ---
 
@@ -36,7 +36,7 @@ memor service install
 memor daemon
 ```
 
-That's it. Every prompt now gets automatic context recall. Open the dashboard to see it working:
+That's it. Every prompt now gets automatic context recall. `memor service install` also starts the dashboard as a background service, so it's already live at http://localhost:8420 (and is recycled whenever you stop/restart/uninstall the service). To run it in the foreground instead:
 
 ```bash
 memor dashboard
@@ -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
@@ -181,10 +182,13 @@ memor install-hook                   Install hook + download model (interactive
 memor daemon                         Auto-ingest + distill (background watcher)
 memor dashboard                      Web dashboard on localhost:8420
 memor version                        Print installed version
-memor service install                Run daemon as background service (launchd/systemd)
-memor service stop                   Stop the background service
-memor service uninstall              Remove the background service
-memor service status                 Check if the service is running
+memor service install                Run daemon + dashboard as background services (launchd/systemd)
+  --no-dashboard                     Install only the daemon
+memor service restart                Restart both services (use after `pipx upgrade`)
+memor service stop                   Stop both background services
+memor service uninstall              Remove both background services
+memor service status                 Show daemon + dashboard status
+  (dashboard port: set MEMOR_DASHBOARD_PORT, default 8420)
 memor query <text>                   Search memories from the CLI
 memor reingest                       Wipe DB and re-ingest everything
 memor reingest --project <name>      Re-ingest only one project
@@ -243,7 +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
 ```
@@ -291,7 +295,7 @@ cd memor-ai
 python3 -m venv .venv && source .venv/bin/activate
 pip install -e ".[dev]"
 
-pytest  # 287 tests
+pytest  # 304 tests
 ```
 
 ---

memor_cli-0.9.0/memor/__init__.py

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

{memor_cli-0.7.1 → memor_cli-0.9.0}/memor/cli.py

@@ -46,15 +46,18 @@ memor — measured memory for coding agents
 GETTING STARTED
   memor install-hook            Install the recall hook + download model
     --agent claude|codex|copilot  Choose your agent (default: claude)
-  memor service install         Start the daemon as a background service
+  memor service install         Start the daemon + dashboard as background services
   memor dashboard               Open the web dashboard at localhost:8420
 
 SERVICE MANAGEMENT
-  memor service install     Install and start as a background service (survives reboots)
-  memor service status      Check if the daemon is running
-  memor service stop        Stop the service (restarts on next login)
-  memor service uninstall   Stop and remove the service completely
+  memor service install     Install/start daemon + dashboard (survives reboots)
+    --no-dashboard            Install only the daemon
+  memor service status      Show daemon + dashboard status
+  memor service restart     Restart both (use after `pipx upgrade`)
+  memor service stop        Stop both services (restart on next login)
+  memor service uninstall   Stop and remove both services completely
   memor daemon              Run the daemon in the foreground (alternative to service)
+  (dashboard port: set MEMOR_DASHBOARD_PORT, default 8420)
 
 QUERYING
   memor query <text>        Search memories from the command line
@@ -218,7 +221,7 @@ def eval_counterfactual_cmd(project: str = typer.Option(...), db: str = "memor.d
         import os
         llm = OpenAICompatLLM(base_url=os.environ.get("OPENAI_BASE_URL", "http://localhost:11434/v1"),
                               api_key=os.environ.get("OPENAI_API_KEY", ""), model=llm_model)
-    summary = run_suite(cases, store=s, embedder=e, llm=llm, k=k)
+    summary = run_suite(cases, store=s, embedder=e, llm=llm, db_path=_db_path(db), k=k)
     typer.echo(json.dumps({k: v for k, v in summary.items() if k != "cases"}, indent=2))
     typer.echo("")
     typer.echo(f"  Win:  {summary['win_count']}/{summary['n_cases']} ({summary['win_pct']}%)")
@@ -638,11 +641,14 @@ def version():
 
 
 @service_app.command("install")
-def service_install():
-    """Install and start the daemon as a background service (survives reboots)."""
+def service_install(
+    no_dashboard: bool = typer.Option(
+        False, "--no-dashboard", help="Install only the daemon, not the dashboard."),
+):
+    """Install and start the daemon + dashboard as background services (survives reboots)."""
     from memor.service import install
     try:
-        typer.echo(install())
+        typer.echo(install(with_dashboard=not no_dashboard))
     except FileNotFoundError as e:
         typer.echo(str(e), err=True)
         raise typer.Exit(1)
@@ -651,9 +657,23 @@ def service_install():
         raise typer.Exit(1)
 
 
+@service_app.command("restart")
+def service_restart():
+    """Restart both services (use after `pipx upgrade` to run the new version)."""
+    from memor.service import restart
+    try:
+        typer.echo(restart())
+    except FileNotFoundError as e:
+        typer.echo(str(e), err=True)
+        raise typer.Exit(1)
+    except Exception as e:
+        typer.echo(f"Failed to restart service: {e}", err=True)
+        raise typer.Exit(1)
+
+
 @service_app.command("uninstall")
 def service_uninstall():
-    """Stop and remove the background service."""
+    """Stop and remove the background services (daemon + dashboard)."""
     from memor.service import uninstall
     typer.echo(uninstall())
 

{memor_cli-0.7.1 → 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-0.7.1 → memor_cli-0.9.0}/memor/eval/counterfactual.py

@@ -13,8 +13,7 @@ import re
 import time
 from dataclasses import dataclass
 
-from memor.types import Artifact, Scope
-from memor.retrieve.retriever import Retriever
+from memor.types import Artifact
 
 
 class Outcome(enum.Enum):
@@ -114,20 +113,38 @@ def parse_verdict_json(raw: str) -> CounterfactualVerdict:
     )
 
 
-def run_case(case: CounterfactualCase, *, store, embedder, llm,
+def run_case(case: CounterfactualCase, *, store, embedder, llm, db_path,
              k: int = 8) -> CounterfactualVerdict:
-    scope = Scope(project=case.scope_project)
-    r = Retriever(store, embedder, k=k)
-    trace = r.query(case.query, scope)
+    """Judge a case using the PRODUCTION recall() path, not a bare Retriever.
+
+    This mirrors what the hook actually injects: same-session exclusion,
+    the 0.3/0.15 score threshold, the per-tier token budget, and 600-char
+    truncation. Without this, the eval over-recalls (self-recall echoes and
+    sub-threshold hits production would never inject), inflating ties and
+    producing same-session "losses" that cannot happen in production.
+    """
+    from memor.query_complexity import route_query, Tier
+    from memor.recall import recall
+
+    tier = route_query(case.query)
+    if tier == Tier.SKIP:
+        return CounterfactualVerdict(
+            outcome=Outcome.TIE,
+            reasoning="Production routes this query to SKIP — no recall would occur",
+            confidence=1.0)
+
+    result = recall(case.query, case.scope_project, db_path, embedder=embedder,
+                    k=tier.k, threshold=0.15, max_tokens=tier.max_tokens,
+                    session_id=case.session_id)
 
-    if not trace.hits:
+    if not result.hit_ids:
         return CounterfactualVerdict(
             outcome=Outcome.TIE,
-            reasoning="No context recalled — nothing to evaluate",
+            reasoning="No context recalled via production path — nothing to evaluate",
             confidence=1.0)
 
-    recalled = "\n\n".join(
-        f"[{h.artifact.kind}] {h.artifact.text}" for h in trace.hits)
+    # Judge exactly what production injects (post-exclusion/threshold/budget/truncation).
+    recalled = result.formatted_context
     holdout = "\n".join(case.holdout_texts)
 
     prompt = COUNTERFACTUAL_PROMPT.format(
@@ -136,11 +153,14 @@ def run_case(case: CounterfactualCase, *, store, embedder, llm,
     return parse_verdict_json(raw)
 
 
-def run_suite(cases: list[CounterfactualCase], *, store, embedder, llm,
+def run_suite(cases: list[CounterfactualCase], *, store, embedder, llm, db_path,
               k: int = 8) -> dict:
     verdicts = []
-    for c in cases:
-        v = run_case(c, store=store, embedder=embedder, llm=llm, k=k)
+    n = len(cases)
+    for i, c in enumerate(cases, 1):
+        print(f"  [{i}/{n}] judging case...", end="", flush=True)
+        v = run_case(c, store=store, embedder=embedder, llm=llm, db_path=db_path, k=k)
+        print(f" {v.outcome.value}", flush=True)
         verdicts.append((c, v))
     verdict_list = [v for _, v in verdicts]
     summary = summarize_verdicts(verdict_list)

{memor_cli-0.7.1 → 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/anthropic.py

@@ -0,0 +1,45 @@
+import time
+
+
+class AnthropicLLM:
+    def __init__(self, model: str = "claude-opus-4-8", api_key: str | None = None,
+                 *, max_retries: int = 8):
+        import anthropic
+
+        # Disable the SDK's own (silent) retries so our visible backoff below
+        # is the single source of waiting — otherwise the SDK sleeps on 429
+        # internally before raising, which looks like a hang.
+        self.client = anthropic.Anthropic(api_key=api_key, max_retries=0)
+        self.model = model
+        self.max_retries = max_retries
+
+    def complete(self, prompt: str, *, max_tokens: int = 1024) -> str:
+        import anthropic
+
+        attempt = 0
+        while True:
+            try:
+                msg = self.client.messages.create(
+                    model=self.model,
+                    max_tokens=max_tokens,
+                    messages=[{"role": "user", "content": prompt}],
+                )
+                return "".join(b.text for b in msg.content if b.type == "text")
+            except anthropic.RateLimitError as e:
+                attempt += 1
+                if attempt > self.max_retries:
+                    raise
+                # Honor Retry-After when present; otherwise wait out the
+                # per-minute token window (tier-1 limits reset each minute).
+                wait = _retry_after_seconds(e) or 60
+                print(f"  rate limited (429); waiting {wait}s then retrying "
+                      f"(attempt {attempt}/{self.max_retries})...", flush=True)
+                time.sleep(wait)
+
+
+def _retry_after_seconds(err) -> float | None:
+    try:
+        val = err.response.headers.get("retry-after")
+        return float(val) if val is not None else None
+    except Exception:
+        return None

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-0.7.1 → 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)