diff --git a/README.md b/README.md
index e92a56d..00321d1 100644
--- a/README.md
+++ b/README.md
@@ -1,3 +1,46 @@
-# ump-recall
+# self-directed-learning
 
-Adaptive Recall sidecar for UMP — multi-channel retrieval (RRF + graph + ACT-R) over UMP memory
\ No newline at end of file
+Self-directed learning framework for Hermes / Krystie — multi-strategy retrieval over UMP memory, with skill detection, decay, and triadic review.
+
+## What this is
+
+A continuous-learning loop for AI agents:
+
+- **Retrieval substrate** — multi-strategy (FTS5 + vector + graph) RRF fusion with ACT-R re-ranking, served as an MCP shim so Hermes/Krystie can use it transparently through their existing UMP tool surface.
+- **Memory hygiene** — nightly decay (per-kind rates, atomic writes, automatic backups), access tracking (frequency + last_accessed_at) feeding the decay and re-ranker.
+- **Knowledge graph** — entity extraction + typed relations + co-occurrence edges; BFS expansion surfaces related URNs even when the direct text doesn't match.
+- **Triadic review** — Claude + Codex (via Hermes OAuth / ChatGPT Pro subscription) as independent judges for claims before they're written to long-term memory.
+- **Cron-driven framework** — skill gap detection, spaced repetition, foraging, reflective journaling, self-measure. All run autonomously and feed back into the retrieval substrate.
+
+## Phases shipped
+
+| Phase | What | Tests |
+|---|---|---|
+| 1 | Sidecar on `:4380` + 3-channel RRF | baseline 0.500 → 1.000 (+50pp) |
+| 2 | Graph channel (typed relations + co-occurrence) | 12 unique top-5 graph-only wins |
+| 3 | ACT-R re-ranker | 84% top-5 retention, 4/20 #1 changes |
+| 4 | Decay script | 20/20 tests, nightly cron |
+| 5 | Access tracking | 28/28 tests, feeds decay + re-ranker |
+| 6 | Co-occurrence edges (in code) | (tests pending) |
+| 7 | Triadic review (judge.py) | Claude + Codex wired |
+
+## Stack
+
+- **Language:** Node.js 18+ (sidecar, MCP shim), Python 3.11+ (decay, framework scripts)
+- **Dependencies:** UMP (`@universalmemoryprotocol/core` 0.1.0), Ollama (`snowflake-arctic-embed2` 1024-dim), Qdrant (`memories_ump` collection), Express, undici, MCP SDK 1.29.
+- **Repo:** `http://100.81.59.99/sami7777/self-directed-learning` (DNS3 Gitea)
+- **Local path:** `/root/ump-recall/` (directory name kept for stability of running processes; only the repo was renamed)
+
+## Run
+
+```bash
+npm start                          # sidecar on :4380
+node src/ump-recall-mcp.js         # MCP shim (stdio)
+python3 scripts/ump_decay.py --apply    # nightly decay
+python3 scripts/judge.py --claim "..."  # Claude + Codex review
+npm test                           # full test suite
+```
+
+## Architecture
+
+See `/root/.hermes/skills/autonomous-ai-agents/agent-self-improvement-framework/` for the full framework spec.
\ No newline at end of file
diff --git a/package.json b/package.json
index 28c5b47..7a343d0 100644
--- a/package.json
+++ b/package.json
@@ -1,7 +1,7 @@
 {
-  "name": "ump-recall",
+  "name": "self-directed-learning",
   "version": "0.1.0",
-  "description": "Adaptive Recall sidecar for UMP — 3-channel RRF (FTS5 + vector + graph) + ACT-R re-ranking",
+  "description": "Self-directed learning framework for Hermes/Krystie: multi-strategy retrieval (3-channel RRF: FTS5 + vector + graph), ACT-R re-ranking, decay, access tracking, and triadic review (Claude + Codex judges).",
   "type": "module",
   "main": "src/server.js",
   "private": true,
diff --git a/scripts/judge.py b/scripts/judge.py
new file mode 100644
index 0000000..776682c
--- /dev/null
+++ b/scripts/judge.py
@@ -0,0 +1,297 @@
+#!/usr/bin/env python3
+"""
+Triadic Self-Learning Review — Phase 7.
+
+Implements the A/B/C triad model for self-improvement:
+  A = Hermes    (the doer)         — produces work, writes memories
+  B = Krystie   (the critic)       — challenges, finds gaps
+  C = Codex + Claude (the judges)  — two external models, structured rubric
+
+Architecture:
+  - judge.py brief-format: structured input to judges, no UMP context
+  - judge.py orchestrates: Claude CLI subprocess + OpenAI API (Codex)
+  - Convergence logic: 2-of-2 agree → accepted; disagree → conflict
+
+Usage:
+  judge.py --claim "I should always pre-cache embeddings before queries"
+  judge.py --memory-urn urn:ump:abc123
+  judge.py --dispute path/to/a-b-exchange.txt
+  judge.py --dry-run --claim "..."
+
+Exit codes:
+  0 = both judges accept (or both reject with strong consensus)
+  1 = judges disagree (needs A/B to resolve or human review)
+  2 = error
+
+Output (stdout, JSON):
+  {
+    "verdict": "accepted" | "rejected" | "conflict",
+    "claude": {"score": 1-5, "reasoning": "..."},
+    "codex":  {"score": 1-5, "reasoning": "..."},
+    "consensus_strength": 0.0-1.0
+  }
+"""
+import argparse
+import json
+import re
+import subprocess
+import sys
+import time
+import urllib.parse
+import urllib.request
+
+JUDGE_PROMPT = """You are a neutral external judge in a triadic self-learning system. The doer and the critic (two AI agents) have produced a claim for your review. They share context, training, and memory. You do NOT — you have no prior context, no memory of them, and your job is to render an independent verdict.
+
+THE CLAIM:
+\"\"\"{claim}\"\"\"
+
+CONTEXT (optional, provided by A or B):
+\"\"\"{context}\"\"\"
+
+RUBRIC — score the claim on each dimension 1-5 (5 = excellent, 1 = poor):
+1. **Truth**: Is the claim factually correct? Check against your knowledge of how systems actually work.
+2. **Specificity**: Is it concrete and actionable, or vague platitude?
+3. **Testability**: Can the claim be verified empirically? Is there a way to confirm or refute it?
+4. **Novelty**: Does this add something genuinely new, or is it a restatement of common practice?
+5. **Harmlessness**: Will following this advice cause damage, leak sensitive data, or create security issues?
+
+Then write a 2-3 sentence verdict: ACCEPT (claim is solid, write to memory), REJECT (claim is wrong, do not write), or REVISE (claim is on the right track but needs changes — suggest them).
+
+Respond ONLY with this exact JSON structure (no markdown, no preamble):
+{{
+  "scores": {{"truth": N, "specificity": N, "testability": N, "novelty": N, "harmlessness": N}},
+  "verdict": "ACCEPT" | "REJECT" | "REVISE",
+  "reasoning": "<2-3 sentences>"
+}}
+"""
+
+
+# --- Judge 1: Claude (via OAuth CLI) ---
+
+def call_claude(claim, context, timeout=120):
+    """Invoke Claude Sonnet via the OAuth CLI as a subprocess.
+    Uses --print for non-interactive mode (output goes to stdout)."""
+    prompt = JUDGE_PROMPT.format(claim=claim, context=context or "(none)")
+    try:
+        proc = subprocess.run(
+            ["/usr/bin/claude", "--model", "sonnet", "--print", prompt],
+            capture_output=True, text=True, timeout=timeout,
+        )
+        if proc.returncode != 0:
+            return {"error": f"claude exit {proc.returncode}: {proc.stderr[:300]}"}
+        return parse_verdict(proc.stdout, "claude")
+    except subprocess.TimeoutExpired:
+        return {"error": "claude timeout"}
+    except Exception as e:
+        return {"error": f"claude exception: {e}"}
+
+
+# --- Judge 2: Codex (via Hermes OAuth — routes to ChatGPT subscription) ---
+
+# Codex is invoked through `hermes chat -m openai-codex/<model>` which uses
+# the OAuth credential already stored via `hermes auth add openai-codex`.
+# This stays on your ChatGPT Pro subscription — no separate API key needed.
+# Verified working 2026-07-12: `hermes chat -m openai-codex/gpt-5.4-mini
+# -q "..."` returns a response.
+DEFAULT_CODEX_MODEL = "openai-codex/gpt-5.4-mini"
+
+
+def call_codex(claim, context, timeout=180, model=None):
+    """Invoke Codex via Hermes OAuth (ChatGPT Pro subscription).
+
+    Routes through `hermes chat -m openai-codex/<model> -q "..." --quiet`
+    which uses the stored credential at `hermes auth list` (openai-codex).
+    No separate token required from the judge script.
+
+    Default model: gpt-5.4-mini (cheap + reasoning-enabled). Override with
+    --codex-model on the CLI or `model=` kwarg here.
+    """
+    use_model = model or DEFAULT_CODEX_MODEL
+    prompt = JUDGE_PROMPT.format(claim=claim, context=context or "(none)")
+    # `-Q` is quiet: suppresses banner/spinner, prints only final response.
+    # `--pass-session-id` is omitted because we don't need to resume.
+    # `--ignore-rules` skips any rule files that might block tool use;
+    # the judge prompt never asks for tools so this is safe.
+    cmd = [
+        "hermes", "chat",
+        "-m", use_model,
+        "-q", prompt,
+        "-Q",  # quiet
+        "--ignore-rules",  # judge prompt doesn't need any rules
+    ]
+    try:
+        proc = subprocess.run(
+            cmd,
+            capture_output=True, text=True, timeout=timeout,
+        )
+        if proc.returncode != 0:
+            return {
+                "_judge": "codex",
+                "error": f"hermes exit {proc.returncode}: {proc.stderr[:400] or proc.stdout[:400]}",
+            }
+        return parse_verdict(proc.stdout, "codex")
+    except subprocess.TimeoutExpired:
+        return {"_judge": "codex", "error": f"hermes timeout after {timeout}s"}
+    except Exception as e:
+        return {"_judge": "codex", "error": f"hermes exception: {e}"}
+
+
+def parse_verdict(text, who):
+    """Parse JSON verdict from a judge's stdout. Tolerant of markdown
+    fences and trailing prose."""
+    text = text.strip()
+    # Strip code fences if present
+    m = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.DOTALL)
+    if m:
+        text = m.group(1)
+    # Find first { ... } block
+    if not text.startswith("{"):
+        start = text.find("{")
+        if start >= 0:
+            text = text[start:]
+    end = text.rfind("}")
+    if end >= 0:
+        text = text[:end + 1]
+    try:
+        v = json.loads(text)
+        v["_judge"] = who
+        return v
+    except json.JSONDecodeError as e:
+        return {"_judge": who, "error": f"parse failed: {e}", "raw": text[:500]}
+
+
+# --- Convergence logic ---
+
+def converge(claude_v, codex_v):
+    """Two-judge convergence. Returns (verdict, consensus_strength, conflicts).
+    verdict ∈ {accepted, rejected, revise, conflict}."""
+    if claude_v.get("error") or codex_v.get("error"):
+        return "conflict", 0.0, ["one_or_more_judges_failed"]
+
+    cv = claude_v.get("verdict", "").upper()
+    xv = codex_v.get("verdict", "").upper()
+
+    # Verdict agreement
+    if cv == xv and cv in ("ACCEPT", "REJECT"):
+        # Compute consensus strength from score similarity
+        c_scores = claude_v.get("scores", {})
+        x_scores = codex_v.get("scores", {})
+        diff = sum(abs(c_scores.get(k, 3) - x_scores.get(k, 3)) for k in ("truth", "specificity", "testability", "novelty", "harmlessness"))
+        # max possible diff = 5 dims × 4 = 20
+        strength = max(0.0, 1.0 - diff / 20.0)
+        return cv.lower(), strength, []
+
+    if cv == xv and cv == "REVISE":
+        # Both want revisions — check if reasoning aligns
+        if claude_v.get("reasoning", "")[:30].lower() == codex_v.get("reasoning", "")[:30].lower():
+            return "revise", 0.7, []
+        return "revise", 0.4, ["revise_suggestions_diverge"]
+
+    # Disagreement
+    return "conflict", 0.0, [f"claude={cv} codex={xv}"]
+
+
+# --- Entry points ---
+
+def main():
+    p = argparse.ArgumentParser(description="Triadic judge — Codex + Claude review a claim")
+    g = p.add_mutually_exclusive_group(required=True)
+    g.add_argument("--claim", help="The claim to judge (string)")
+    g.add_argument("--memory-urn", help="Fetch a UMP memory by urn and judge it")
+    g.add_argument("--dispute", help="Path to a file containing A/B exchange text")
+    g.add_argument("--dispute-stdin", action="store_true", help="Read dispute from stdin")
+    p.add_argument("--context", default="", help="Optional context for the claim")
+    p.add_argument("--dry-run", action="store_true", help="Print prompt, don't call judges")
+    p.add_argument("--timeout", type=int, default=120, help="Per-judge timeout seconds")
+    p.add_argument("--skip-claude", action="store_true")
+    p.add_argument("--skip-codex", action="store_true")
+    p.add_argument("--codex-model", default=DEFAULT_CODEX_MODEL,
+                   help=f"Codex model name (default: {DEFAULT_CODEX_MODEL})")
+    args = p.parse_args()
+
+    claim = args.claim or ""
+    context = args.context or ""
+
+    if args.memory_urn:
+        claim, context = fetch_memory(args.memory_urn)
+    elif args.dispute:
+        text = open(args.dispute).read()
+        claim, context = parse_dispute(text)
+    elif args.dispute_stdin:
+        text = sys.stdin.read()
+        claim, context = parse_dispute(text)
+
+    if args.dry_run:
+        print("=== DRY RUN — would call judges with this prompt ===")
+        print(JUDGE_PROMPT.format(claim=claim[:500], context=context[:500]))
+        return
+
+    print(f"Claim: {claim[:120]}{'...' if len(claim) > 120 else ''}")
+    print(f"Context: {context[:80] if context else '(none)'}")
+    print()
+
+    claude_v = {}
+    codex_v = {}
+    if not args.skip_claude:
+        print("Calling Claude Sonnet...", end=" ", flush=True)
+        t0 = time.time()
+        claude_v = call_claude(claim, context, timeout=args.timeout)
+        print(f"({time.time()-t0:.1f}s)")
+    if not args.skip_codex:
+        print(f"Calling Codex ({args.codex_model} via Hermes OAuth)...", end=" ", flush=True)
+        t0 = time.time()
+        codex_v = call_codex(claim, context, timeout=args.timeout, model=args.codex_model)
+        print(f"({time.time()-t0:.1f}s)")
+    print()
+
+    verdict, strength, conflicts = converge(claude_v, codex_v)
+    result = {
+        "verdict": verdict,
+        "consensus_strength": round(strength, 3),
+        "conflicts": conflicts,
+        "claude": claude_v,
+        "codex": codex_v,
+        "claim_preview": claim[:200],
+    }
+    print(json.dumps(result, indent=2))
+
+    sys.exit(0 if verdict in ("accepted", "rejected", "revise") else 1)
+
+
+def fetch_memory(urn):
+    """Fetch a UMP memory and format as claim + context."""
+    try:
+        req = urllib.request.Request(
+            f"http://127.0.0.1:4317/ump/memory/{urllib.parse.quote(urn, safe='')}",
+        )
+        with urllib.request.urlopen(req, timeout=10) as r:
+            data = json.loads(r.read())
+        rec = data.get("record") or data
+        body = rec.get("body", {})
+        subject = body.get("subject", "")
+        text = body.get("text", "")
+        claim = f"{subject}\n\n{text}" if subject else text
+        context = f"kind={rec.get('kind')} topic={rec.get('scope',{}).get('project','')}"
+        return claim, context
+    except Exception as e:
+        return f"(failed to fetch {urn}: {e})", ""
+
+
+def parse_dispute(text):
+    """Parse a structured A/B exchange. Expects lines starting with 'A:' or 'B:'."""
+    a_lines, b_lines = [], []
+    for line in text.splitlines():
+        if line.startswith("A:"):
+            a_lines.append(line[2:].strip())
+        elif line.startswith("B:"):
+            b_lines.append(line[2:].strip())
+    a_text = "\n".join(a_lines).strip()
+    b_text = "\n".join(b_lines).strip()
+    # The claim under dispute is the disagreement. We synthesize:
+    claim = a_text if a_text else b_text
+    context = f"A says:\n{a_text}\n\nB says:\n{b_text}" if a_text and b_text else ""
+    return claim, context
+
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file
diff --git a/src/ump-recall-mcp.js b/src/ump-recall-mcp.js
index 870e53e..5377e5e 100644
--- a/src/ump-recall-mcp.js
+++ b/src/ump-recall-mcp.js
@@ -130,12 +130,14 @@ function spawnUmpChild() {
   umpChild.on("exit", (code) => {
     LOG(`UMP subprocess exited code=${code}`);
     umpChild = null;
+    umpChildInitPromise = null; // force re-initialize on next call
   });
   return umpChild;
 }
 
 let nextReqId = 1;
 const pendingRequests = new Map();
+let umpServerInfo = null; // populated by initialize handshake
 
 function setupUmpStdio(child) {
   let buf = "";
@@ -161,13 +163,78 @@ function setupUmpStdio(child) {
   });
 }
 
-function umpCall(method, params) {
-  return new Promise((resolve, reject) => {
-    const child = spawnUmpChild();
-    if (!umpChildInitPromise) {
-      setupUmpStdio(child);
-      umpChildInitPromise = Promise.resolve();
+// Send the MCP `initialize` handshake to the canonical UMP child. This MUST
+// happen before any `tools/call` — the MCP spec requires it, and the SDK
+// rejects tools/call from a not-yet-initialized client. We do this exactly
+// once per child lifetime, lazily on the first umpCall().
+async function ensureUmpInitialized(child) {
+  if (umpChildInitPromise) return umpChildInitPromise;
+
+  setupUmpStdio(child);
+
+  // Use the modern MCP protocol version. The SDK accepts whatever the
+  // server negotiates back; we just need to send *something* valid.
+  const PROTOCOL_VERSION = "2024-11-05";
+  const initPromise = new Promise((resolve, reject) => {
+    const id = nextReqId++;
+    pendingRequests.set(id, { resolve, reject });
+    const msg = JSON.stringify({
+      jsonrpc: "2.0",
+      id,
+      method: "initialize",
+      params: {
+        protocolVersion: PROTOCOL_VERSION,
+        capabilities: {},
+        clientInfo: { name: "ump-recall-mcp-shim", version: "0.1.0" },
+      },
+    });
+    const timer = setTimeout(() => {
+      if (pendingRequests.has(id)) {
+        pendingRequests.delete(id);
+        reject(new Error("UMP initialize timed out after 10s"));
+      }
+    }, 10000);
+    try {
+      child.stdin.write(msg + "\n");
+    } catch (e) {
+      clearTimeout(timer);
+      pendingRequests.delete(id);
+      reject(e);
     }
+  }).then((result) => {
+    umpServerInfo = result?.serverInfo || null;
+    LOG(`UMP initialized: server=${umpServerInfo?.name} v${umpServerInfo?.version}`);
+    // Per MCP spec, after initialize the client must send `notifications/initialized`.
+    // The server may not reply to notifications (fire-and-forget JSON-RPC).
+    try {
+      child.stdin.write(JSON.stringify({
+        jsonrpc: "2.0",
+        method: "notifications/initialized",
+        params: {},
+      }) + "\n");
+    } catch (e) {
+      LOG("notifications/initialized write failed (non-fatal):", e.message);
+    }
+    return result;
+  });
+
+  umpChildInitPromise = initPromise;
+  return initPromise;
+}
+
+async function umpCall(method, params) {
+  const child = spawnUmpChild();
+  // Ensure the child has gone through the MCP initialize handshake.
+  // This lazily sets up stdio listeners on first call and caches the
+  // serverInfo so we don't re-initialize per request.
+  try {
+    await ensureUmpInitialized(child);
+  } catch (e) {
+    // If initialize fails (e.g. UMP not installed), propagate so callers
+    // get a real error instead of a silent timeout.
+    throw new Error(`UMP initialize failed: ${e.message}`);
+  }
+  return new Promise((resolve, reject) => {
     const id = nextReqId++;
     pendingRequests.set(id, { resolve, reject });
     const msg = JSON.stringify({ jsonrpc: "2.0", id, method, params: params || {} });