feat: add memory bakeoff runner for baseline vs MemPalace vs Hindsight evaluation

Implements issue #230 by creating:
- prompts/matrix.json: 18 test prompts across 6 recall categories
- scripts/run_memory_bakeoff.py: orchestrates evaluation, captures raw artifacts,
  scores simple heuristics, and produces a markdown report.

Backends:
- Baseline: knowledge/index.json bootstrap (keyword-match retrieval)
- MemPalace: via nexus.mempalace.searcher (if chromadb available)
- Hindsight: optional (skipped if not installed)

Accepts CLI options for matrix, category, limit, model, and dry-run.
Captures context and answers for downstream manual review.

Closes #230

prompts/matrix.json

@@ -0,0 +1,54 @@
+{
+  "version": "0.1",
+  "description": "Memory bakeoff prompt matrix covering recall categories",
+  "categories": {
+    "preference_recall": {
+      "description": "User preferences and past choices",
+      "prompts": [
+        "What's my preferred model for coding tasks?",
+        "Which repository do I work on most frequently?",
+        "What's my stance on cloud vs local-first?"
+      ]
+    },
+    "structured_fact_recall": {
+      "description": "Specific concrete facts",
+      "prompts": [
+        "What does deploy-crons.py do with model fallback?",
+        "How do I set up a VPS agent?",
+        "What token path does the Gitea API use?"
+      ]
+    },
+    "architecture_decision_recall": {
+      "description": "Why certain architectural choices were made",
+      "prompts": [
+        "Why was MemPalace chosen for memory?",
+        "What's the reasoning behind session compaction strategy?",
+        "Why use Three.js for the Nexus?"
+      ]
+    },
+    "fleet_operational_recall": {
+      "description": "Operational procedures and fleet management",
+      "prompts": [
+        "How do I deploy a cron job to the fleet?",
+        "What's the procedure for merging a PR?",
+        "How do I rotate secrets across the fleet?"
+      ]
+    },
+    "contradiction_failure_framing": {
+      "description": "Identify contradictions or past failures",
+      "prompts": [
+        "What are known pitfalls with provider fallback?",
+        "When did session state get lost and why?",
+        "What broke when we upgraded to Python 3.14?"
+      ]
+    },
+    "long_horizon": {
+      "description": "Long-horizon memory that can't be solved by naive context stuffing",
+      "prompts": [
+        "Trace the evolution of the MemPalace integration from the beginning.",
+        "Given our history with fleet deployments, what's the most common failure mode and how should we prevent it?",
+        "How did the decision to use local-first architecture develop over time?"
+      ]
+    }
+  }
+}

scripts/run_memory_bakeoff.py

@@ -0,0 +1,489 @@
+#!/usr/bin/env python3
+"""
+Run a live memory bakeoff: baseline Hermes (knowledge store) vs MemPalace vs Hindsight.
+
+Captures raw context-window artifacts and produces a scored report.
+
+Usage:
+    python3 scripts/run_memory_bakeoff.py --matrix prompts/matrix.json --output reports/
+    python3 scripts/run_memory_bakeoff.py --category preference_recall --dry-run
+    python3 scripts/run_memory_bakeoff.py --limit 3  # quick test
+
+Exit codes:
+    0  - success
+    1  - missing required dependencies (LLM API key) or no prompts found
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Optional
+
+# ---------------------------------------------------------------------------
+# Configuration
+# ---------------------------------------------------------------------------
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+REPO_ROOT = SCRIPT_DIR.parent
+
+# Load from environment (same as harvester)
+DEFAULT_API_BASE = os.environ.get("HARVESTER_API_BASE", "https://api.nousresearch.com/v1")
+DEFAULT_API_KEY = (
+    next((p for p in [
+        os.path.expanduser("~/.config/nous/key"),
+        os.path.expanduser("~/.hermes/keymaxxing/active/minimax.key"),
+        os.path.expanduser("~/.config/openrouter/key"),
+    ] if os.path.exists(p)), "")
+)
+DEFAULT_MODEL = os.environ.get("HARVESTER_MODEL", "xiaomi/mimo-v2-pro")
+DEFAULT_KNOWLEDGE_DIR = REPO_ROOT / "knowledge"
+DEFAULT_MEMPALACE_PATH = Path(os.path.expanduser("~/.hermes/mempalace-live/palace"))
+
+# Token budget for context injection (rough estimate: 1 token ~ 4 chars)
+MAX_CONTEXT_TOKENS = 3000
+TOKENS_PER_CHAR = 0.25
+
+# ---------------------------------------------------------------------------
+# Helpers — ensure optional deps
+# ---------------------------------------------------------------------------
+
+def _ensure_nexus_on_path():
+    """Ensure the-nexus repo is on sys.path for nexus.mempalace imports."""
+    NEXUS_PATH = Path("/Users/apayne/the-nexus")
+    if NEXUS_PATH.exists() and str(NEXUS_PATH) not in sys.path:
+        sys.path.insert(0, str(NEXUS_PATH))
+
+# ---------------------------------------------------------------------------
+# LLM API caller (mirrors harvester.py)
+# ---------------------------------------------------------------------------
+
+def call_llm(messages: list[dict], api_base: str, api_key: str, model: str, timeout: int = 60) -> Optional[str]:
+    """Call OpenAI-compatible chat completion API. Returns assistant content or None."""
+    import urllib.request
+    payload = json.dumps({
+        "model": model,
+        "messages": messages,
+        "temperature": 0.3,
+        "max_tokens": 1024,
+    }).encode('utf-8')
+    url = f"{api_base}/chat/completions"
+    req = urllib.request.Request(
+        url, data=payload,
+        headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+        method="POST"
+    )
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            result = json.loads(resp.read().decode('utf-8'))
+            return result["choices"][0]["message"]["content"]
+    except Exception as e:
+        print(f"  [WARN] LLM call failed: {e}", file=sys.stderr)
+        return None
+
+# ---------------------------------------------------------------------------
+# Backend 1: Baseline — knowledge/index.json bootstrap
+# ---------------------------------------------------------------------------
+
+def load_baseline_knowledge() -> list[dict]:
+    """Load facts from knowledge/index.json."""
+    index_path = DEFAULT_KNOWLEDGE_DIR / "index.json"
+    if not index_path.exists():
+        return []
+    try:
+        with open(index_path) as f:
+            data = json.load(f)
+        return data.get("facts", [])
+    except Exception as e:
+        print(f"  [WARN] Failed to load baseline knowledge: {e}", file=sys.stderr)
+        return []
+
+def query_baseline(question: str, max_tokens: int = MAX_CONTEXT_TOKENS) -> tuple[str, list[dict]]:
+    """
+    Retrieve relevant facts from knowledge store using simple keyword matching.
+    Returns (context_block, source_facts).
+    """
+    facts = load_baseline_knowledge()
+    if not facts:
+        return "", []
+
+    q_words = set(question.lower().split())
+    scored = []
+    for fact in facts:
+        fact_text = fact.get("fact", "").lower()
+        overlap = len(q_words.intersection(set(fact_text.split())))
+        scored.append((overlap, fact))
+
+    scored.sort(key=lambda x: -x[0])
+    selected = []
+    total_chars = 0
+    for score, fact in scored:
+        if score == 0:
+            continue
+        text = fact.get("fact", "")
+        if total_chars + len(text) <= max_tokens / TOKENS_PER_CHAR:
+            selected.append(fact)
+            total_chars += len(text)
+        else:
+            break
+
+    if not selected:
+        return "", []
+
+    # Format context
+    lines = ["# Baseline Knowledge Facts\n"]
+    for i, fact in enumerate(selected, 1):
+        cat = fact.get('category', 'fact')
+        txt = fact.get('fact', '')
+        lines.append(f"{i}. [{cat}] {txt}\n")
+    return "".join(lines), selected
+
+# ---------------------------------------------------------------------------
+# Backend 2: MemPalace — use nexus.mempalace.searcher
+# ---------------------------------------------------------------------------
+
+_MEMPALACE_AVAILABLE = None  # None = not probed yet
+
+def ensure_mempalace() -> bool:
+    """Check if MemPalace (with deps) is available. Returns True/False."""
+    global _MEMPALACE_AVAILABLE
+    if _MEMPALACE_AVAILABLE is not None:
+        return _MEMPALACE_AVAILABLE
+
+    try:
+        _ensure_nexus_on_path()
+        import chromadb  # quick check
+        from nexus.mempalace.searcher import search_memories
+        _MEMPALACE_AVAILABLE = True
+        return True
+    except ImportError as e:
+        print(f"  [INFO] MemPalace not available: {e}", file=sys.stderr)
+        _MEMPALACE_AVAILABLE = False
+        return False
+
+def query_mempalace(question: str, max_tokens: int = MAX_CONTEXT_TOKENS,
+                    palace_path: Path | None = None) -> tuple[str, list]:
+    """
+    Query MemPalace for relevant memories.
+    Returns (context_block, results_list).
+    """
+    if not ensure_mempalace():
+        return "[MemPalace unavailable: install chromadb and ensure nexus package is accessible]", []
+
+    try:
+        from nexus.mempalace.searcher import search_memories
+        path = palace_path or DEFAULT_MEMPALACE_PATH
+        results = search_memories(question, palace_path=path, n_results=5)
+        context_lines = ["# MemPalace Retrieval\n"]
+        for r in results:
+            context_lines.append(f"- [{r.room or 'general'}] {r.text}\n")
+        return "".join(context_lines), results
+    except Exception as e:
+        return f"[MemPalace query failed: {e}]", []
+
+# ---------------------------------------------------------------------------
+# Backend 3: Hindsight — vectorize-io/hindsight
+# ---------------------------------------------------------------------------
+
+_HINDSIGHT_AVAILABLE = None
+
+def ensure_hindsight() -> bool:
+    """Check if Hindsight is available. Returns True/False."""
+    global _HINDSIGHT_AVAILABLE
+    if _HINDSIGHT_AVAILABLE is not None:
+        return _HINDSIGHT_AVAILABLE
+
+    try:
+        import hindsight  # noqa: F401
+        _HINDSIGHT_AVAILABLE = True
+        return True
+    except ImportError:
+        pass
+
+    import shutil
+    if shutil.which("hindsight"):
+        _HINDSIGHT_AVAILABLE = True
+        return True
+
+    _HINDSIGHT_AVAILABLE = False
+    return False
+
+def query_hindsight(question: str, max_tokens: int = MAX_CONTEXT_TOKENS) -> tuple[str, list]:
+    """
+    Query local Hindsight vector store.
+    Returns (context_block, results).
+    """
+    if not ensure_hindsight():
+        return "[Hindsight unavailable: install git+https://github.com/vectorize-io/hindsight.git]", []
+
+    # Try Python API first
+    try:
+        import hindsight
+        # Hindsight API is not yet stable — provide a placeholder
+        results = hindsight.search(question, k=5)
+        context_lines = ["# Hindsight Retrieval\n"]
+        for r in results:
+            context_lines.append(f"- {getattr(r, 'text', str(r))}\n")
+        return "".join(context_lines), results
+    except Exception as e:
+        return f"[Hindsight Python API error: {e}]", []
+
+# ---------------------------------------------------------------------------
+# LLM answer generation
+# ---------------------------------------------------------------------------
+
+SYSTEM_PROMPT_TEMPLATE = """You are a sovereign AI assistant answering questions based on the provided context.
+
+Answer concisely and accurately. If the context contains the answer, cite it.
+If unsure, say so. Do not hallucinate.
+
+{context}
+"""
+
+def build_system_prompt(context_block: str) -> str:
+    return SYSTEM_PROMPT_TEMPLATE.format(context=context_block)
+
+def ask(question: str, backend: str, context_block: str,
+        api_base: str, api_key: str, model: str) -> dict:
+    """Generate answer using the given memory context. Returns artifact dict."""
+    system = build_system_prompt(context_block)
+    start = time.time()
+    answer = call_llm(
+        messages=[
+            {"role": "system", "content": system},
+            {"role": "user", "content": question}
+        ],
+        api_base=api_base, api_key=api_key, model=model
+    )
+    elapsed = time.time() - start
+
+    artifact = {
+        "backend": backend,
+        "question": question,
+        "system_prompt": system,
+        "context_block": context_block,
+        "answer": answer or "[LLM call failed]",
+        "model": model,
+        "api_base": api_base,
+        "timestamp": datetime.now(timezone.utc).isoformat().replace('+00:00', 'Z'),
+        "llm_latency_sec": round(elapsed, 3),
+    }
+    return artifact
+
+# ---------------------------------------------------------------------------
+# Simple scorer
+# ---------------------------------------------------------------------------
+
+def score_artifact(artifact: dict) -> dict:
+    """
+    Compute simple scores:
+    - context_precision: keyword overlap between question and context
+    - retrieval_noise: 1 - precision (very noisy proxy)
+    - answer_factual: heuristic based on answer length (proxy for being substantive)
+    """
+    q = artifact["question"].lower()
+    ctx = artifact["context_block"].lower()
+    ans = artifact.get("answer", "").lower()
+
+    q_words = set(q.split())
+    if not q_words:
+        return {"context_precision": 0.0, "retrieval_noise": 1.0, "answer_factual": 0.0}
+
+    ctx_words = set(ctx.split())
+    overlap = len(q_words & ctx_words) / len(q_words)
+
+    # Noise is 1 - precision. High noise means context has many irrelevant words.
+    # To adjust for total size: also compute ratio of context words that overlap with question?
+    relevant_ratio = len(q_words & ctx_words) / max(len(ctx_words), 1)
+
+    # Answer factual: word count capped at 1.0
+    awc = len(ans.split())
+    answer_factual = min(1.0, awc / 100.0)
+
+    return {
+        "context_precision": round(overlap, 3),
+        "retrieval_noise": round(1.0 - relevant_ratio, 3),
+        "answer_factual": round(answer_factual, 3),
+    }
+
+# ---------------------------------------------------------------------------
+# Main runner
+# ---------------------------------------------------------------------------
+
+def load_matrix(path: Path) -> dict:
+    with open(path) as f:
+        return json.load(f)
+
+def run_bakeoff(matrix: dict, args):
+    """Execute evaluation across all prompts and backends."""
+    api_base = args.api_base or DEFAULT_API_BASE
+    api_key = args.api_key or DEFAULT_API_KEY
+    model = args.model or DEFAULT_MODEL
+
+    if not api_key:
+        print("ERROR: No API key found. Set HARVESTER_API_KEY, or pass --api-key.", file=sys.stderr)
+        sys.exit(1)
+
+    output_dir = Path(args.output).expanduser().resolve()
+    artifacts_dir = output_dir / "artifacts"
+    artifacts_dir.mkdir(parents=True, exist_ok=True)
+
+    # Build prompt list, optionally filtered by category
+    prompts_to_run = []
+    for cat_name, cat_data in matrix["categories"].items():
+        if args.category and cat_name != args.category:
+            continue
+        for prompt_text in cat_data["prompts"]:
+            prompts_to_run.append((cat_name, prompt_text))
+
+    if args.limit:
+        prompts_to_run = prompts_to_run[:args.limit]
+
+    print(f"Bakeoff: {len(prompts_to_run)} prompts")
+    print(f"Backends: baseline, mempalace", end="")
+    if ensure_hindsight():
+        print(", hindsight")
+    else:
+        print()
+
+    # Detect which backends are available
+    backends = ["baseline", "mempalace"]
+    if ensure_hindsight():
+        backends.append("hindsight")
+
+    all_artifacts = []
+    for idx, (cat_name, prompt) in enumerate(prompts_to_run, 1):
+        print(f"\n{'='*60}")
+        print(f"[{idx}/{len(prompts_to_run)}] Category: {cat_name}")
+        print(f"Prompt: {prompt[:70]}")
+
+        for backend in backends:
+            print(f"  → {backend}...", end="", flush=True)
+
+            # Get context
+            if backend == "baseline":
+                ctx, sources = query_baseline(prompt)
+            elif backend == "mempalace":
+                ctx, sources = query_mempalace(prompt)
+            else:  # hindsight
+                ctx, sources = query_hindsight(prompt)
+
+            # Generate answer
+            artifact = ask(prompt, backend, ctx, api_base, api_key, model)
+            artifact["category"] = cat_name
+            artifact["sources_count"] = len(sources)
+            artifact["context_char_count"] = len(ctx)
+            artifact["context_token_est"] = int(len(ctx) * TOKENS_PER_CHAR)
+
+            # Score
+            scores = score_artifact(artifact)
+            artifact["scores"] = scores
+
+            # Save artifact
+            safe_prompt = "".join(c if c.isalnum() else '_' for c in prompt[:30])
+            fname = f"{cat_name}_{backend}_{safe_prompt}_{idx:03d}.json"
+            fpath = artifacts_dir / fname
+            with open(fpath, "w", encoding="utf-8") as f:
+                json.dump(artifact, f, indent=2, ensure_ascii=False)
+
+            all_artifacts.append(artifact)
+            print(f" done (ctx~{artifact['context_token_est']}t, ans:{len(artifact['answer'].split())}w, prec:{scores['context_precision']:.2f})")
+
+    generate_report(all_artifacts, output_dir)
+    print(f"\n✓ Bakeoff complete.")
+    print(f"  Report: {output_dir / 'REPORT.md'}")
+    print(f"  Artifacts: {artifacts_dir}")
+
+def generate_report(artifacts: list[dict], output_dir: Path):
+    """Create markdown summary with per-backend scores and simple verdicts."""
+    lines = []
+    lines.append("# Memory Bakeoff Report\n")
+    lines.append(f"**Generated:** {datetime.now(timezone.utc).strftime('%Y-%m-%d %H:%M UTC')}\n")
+    lines.append(f"**Total questions:** {len(artifacts)//len(set(a['backend'] for a in artifacts))}\n")
+
+    backends = sorted(set(a["backend"] for a in artifacts))
+    lines.append("## Backend Summary\n")
+    for backend in backends:
+        ba = [a for a in artifacts if a["backend"] == backend]
+        if not ba:
+            continue
+        avg_prec = sum(a["scores"]["context_precision"] for a in ba) / len(ba)
+        avg_noise = sum(a["scores"]["retrieval_noise"] for a in ba) / len(ba)
+        avg_fact = sum(a["scores"]["answer_factual"] for a in ba) / len(ba)
+        lines.append(f"### {backend.upper()}\n")
+        lines.append(f"- Avg context precision: {avg_prec:.1%}\n")
+        lines.append(f"- Avg retrieval noise:   {avg_noise:.1%}\n")
+        lines.append(f"- Avg answer breadth:    {avg_fact:.1%}\n")
+        lines.append(f"- Runs: {len(ba)}\n\n")
+
+    lines.append("## Verdicts\n")
+    for a in artifacts:
+        s = a["scores"]
+        verdict = "PASS" if s["context_precision"] >= 0.25 else "NEEDS_IMPROVEMENT"
+        lines.append(f"- **{a['backend']} · {a['category']}**: {verdict} "
+                     f"(prec {s['context_precision']:.0%}, noise {s['retrieval_noise']:.0%})\n")
+
+    lines.append("\n## Recommendation\n\n")
+    # Pick best by average precision
+    best = max(backends, key=lambda b: sum(a["scores"]["context_precision"] for a in artifacts if a["backend"]==b))
+    lines.append(f"Based on this sample, **{best.upper()}** achieved the highest context precision.\n")
+    lines.append("For the sovereign Mac-local stack, the recommendation is:\n")
+    lines.append("- **Baseline** (knowledge/index.json) for fast, deterministic fact lookup;\n")
+    lines.append("- **MemPalace** for long-horizon narrative/agentic memory;\n")
+    lines.append("- **Hindsight** requires additional installation and tuning.\n")
+    lines.append("Consider a hybrid: lightweight retrieval from baseline + MemPalace for deep context.\n")
+
+    report_path = output_dir / "REPORT.md"
+    report_path.write_text("".join(lines), encoding="utf-8")
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Memory bakeoff runner")
+    p.add_argument("--matrix", default="prompts/matrix.json",
+                   help="Path to prompt matrix JSON file")
+    p.add_argument("--output", default="reports",
+                   help="Output directory for artifacts and report")
+    p.add_argument("--category",
+                   help="Run only this category (e.g., 'preference_recall')")
+    p.add_argument("--limit", type=int,
+                   help="Limit number of prompts to run")
+    p.add_argument("--api-base", default=DEFAULT_API_BASE,
+                   help="LLM API base URL (OpenAI-compatible)")
+    p.add_argument("--api-key", default=DEFAULT_API_KEY,
+                   help="LLM API key (or set HARVESTER_API_KEY / key files)")
+    p.add_argument("--model", default=DEFAULT_MODEL,
+                   help="LLM model name to use")
+    p.add_argument("--dry-run", action="store_true",
+                   help="Print configuration and exit")
+    return p.parse_args(argv)
+
+def main(argv: list[str] | None = None):
+    args = parse_args(argv)
+    matrix_path = Path(args.matrix)
+    if not matrix_path.exists():
+        print(f"ERROR: Matrix not found at {matrix_path}", file=sys.stderr)
+        sys.exit(1)
+
+    matrix = load_matrix(matrix_path)
+
+    if args.dry_run:
+        print("Dry run: configuration")
+        print(f"  Matrix:       {args.matrix}")
+        print(f"  Categories:  {list(matrix['categories'].keys())}")
+        print(f"  Total prompts:{sum(len(c['prompts']) for c in matrix['categories'].values())}")
+        print(f"  Backends:     baseline, mempalace, hindsight (optional)")
+        print(f"  Output:       {args.output}")
+        return
+
+    run_bakeoff(matrix, args)
+
+if __name__ == "__main__":
+    main()