feat: add evidence-bound compiled user profile

.planning/COMPILED-USER-PROFILE-V1.md

@@ -0,0 +1,47 @@
+<!-- doc-head: compiled user profile v1 implementation contract -->
+# Compiled User Profile V1
+# Covers: evidence input, weekly GLM map/reduce, deterministic projection, and injection.
+# Key terms: verbatim_memories, exact supports, user.md, SessionStart, feature flag.
+# Read when: changing profile extraction, lifecycle, scheduling, rendering, or rollout.
+# Status: implemented on an isolated feature branch; integration and activation pending.
+<!-- /doc-head -->
+
+## Contract
+
+MemoryMaster's SQLite database remains authoritative. The generated
+`~/.memorymaster/projections/user.md` is a bounded, disposable view of stable
+user facts and preferences; it is never parsed back into claims and contains
+facts, not agent instructions.
+
+- Input is incremental sanitized `verbatim_memories`: user turns are evidence;
+  the preceding assistant turn is bounded context only.
+- GLM map output proposes allowlisted facts with exact verbatim row IDs. GLM
+  reduce output must partition every candidate into add, reinforce, replace, or
+  ignore. Unknown IDs, sensitive content, malformed JSON, and instruction-shaped
+  text fail closed.
+- New or replacement facts require support from at least two independent
+  sessions. SQLite records exact row IDs, session IDs, message hashes, and dates.
+- Stable facts survive silence. Preferences expire after 90 unsupported days.
+- A deterministic renderer writes at most 40 facts within an 800-token budget.
+- The existing Dreaming task runs at most three map calls per invocation and
+  resumes from durable watermarks. `MEMORYMASTER_COMPILED_PROFILE=1` enables it;
+  the default is off.
+- SessionStart injects only a bounded file carrying MemoryMaster's generated
+  marker. Hand-written or oversized files are ignored.
+
+## Operator commands
+
+```powershell
+python -m memorymaster.profile status --db .\memorymaster.db
+python -m memorymaster.profile run --db .\memorymaster.db --workspace . --force
+```
+
+## Acceptance evidence
+
+- Focused engine tests cover incremental extraction, strict support validation,
+  resumability, exact support lineage, independent-session gating, preference
+  expiry, stable-fact retention, and deterministic budget bounds.
+- Surface tests cover feature-off scheduling, fail-closed enabled scheduling,
+  generated-only SessionStart loading, and CLI status/help.
+- Activation, a public package release, and historical transcript bootstrap are
+  separate operator actions; this implementation does not perform them.

DOCS-MAP.md

@@ -1,17 +1,18 @@
-<!-- doc-head: canonical documentation map; P5 integrated, PPR-7 active -->
+<!-- doc-head: canonical map; PPR-7 integrated, compiled profile in development -->
 # DOCS-MAP - memorymaster
 # Covers: trust verdicts and replacements for every canonical documentation surface.
-# Key terms: CURRENT, SUPERSEDED, ABANDONED, GENERATED, roadmap, paper radar, ADR.
+# Key terms: CURRENT, SUPERSEDED, compiled user profile, roadmap, ADR.
 # Read when: locating authoritative project documentation before reading doc bodies.
-# Updated: 2026-08-12 after PR #189 merge, governed runtime deploy, and PPR-7 authorization.
-# Rule: PPR-7 is isolated and opt-in; public release and automatic recall remain separately gated.
+# Updated: 2026-08-12 for the evidence-bound compiled user-profile implementation.
+# Rule: generated profiles are disposable projections; MemoryMaster remains authority.
 <!-- /doc-head -->
 
 | File | Verdict | Last change | Reason |
 |---|---|---|---|
 | CHANGELOG.md | CURRENT | 2026-08-04 | Public release history; v4.6.0 records governed universal capture, measured quality changes, security evidence, and known follow-ups. |
 | ROADMAP.md | CURRENT | 2026-08-12 | Sole authoritative roadmap; P5 is merged and locally deployed, while isolated PPR-7 implementation is active. |
 | .planning/GRAPH-OBSERVATIONS-V1.md | CURRENT | 2026-08-12 | Bounded PPR-7 implementation ledger for deterministic supported components, governed synthesis, lifecycle, opt-in recall, and verification. |
+| .planning/COMPILED-USER-PROFILE-V1.md | CURRENT | 2026-08-12 | Implementation contract for the zero-curation, evidence-bound user profile projection and weekly GLM map/reduce job. |
 | .planning/PAPER-RADAR-REVIEW-2026-08-08.md | CURRENT | 2026-08-08 | Primary-paper ledger covers 57-paper triage, 18 deep reviews, exact MemoryMaster gaps, and ordered PPR-1 through PPR-6 decisions subordinate to ROADMAP.md. |
 | .planning/PAPER-RESEARCH-IMPLEMENTATION-2026-08-08.md | CURRENT | 2026-08-08 | Executable status ledger for PPR-1 through PPR-6; records acceptance criteria and evidence without competing with ROADMAP.md. |
 | .planning/HERMES-SCOPE-SKILLS-INTEGRATION-2026-08-07.md | CURRENT | 2026-08-12 | Executable Tencent-derived ledger; repair5 passed and PR #189 is open without merge authority. |

ROADMAP.md

@@ -1,10 +1,10 @@
-<!-- doc-head: sole authoritative roadmap; P5 merged/deployed, PPR-7 active -->
+<!-- doc-head: sole roadmap; PPR-7 integrated, compiled profile in development -->
 # MemoryMaster roadmap
-# Covers: post-v4.6 sequence, Tencent-derived work, paper research, and deferrals.
-# Key terms: Hermes, governed skills, paper radar, temporal projection, sustainability.
+# Covers: post-v4.6 sequence, governed observations, compiled profile, and deferrals.
+# Key terms: Hermes, governed skills, graph observations, user profile, sustainability.
 # Read when: choosing release scope, accepting a feature, or checking deferrals.
 # Authority: sole roadmap; planning ledgers implement it and never replace it.
-# Safety: SQLite authority and steward promotion remain fixed; PPR-7 stays additive, opt-in, and unreleased pending its own evidence.
+# Safety: SQLite remains authority; generated user.md is disposable and feature-off by default.
 <!-- /doc-head -->
 
 ## Shipped in v4.6.0
@@ -68,6 +68,10 @@ its governed-claims authority:
   opt-in; its separate PR/CI, feature-off deployment, and rollout evidence are
   still required before activation. No public release or automatic
   ordinary-recall inclusion is implied.
+- The compiled user profile is being implemented as a disposable projection of
+  sanitized transcript evidence. Weekly bounded GLM map/reduce work stores
+  exact supports in SQLite; deterministic rendering produces `user.md` for
+  SessionStart injection. The feature remains off until explicitly enabled.
 - The invalid earlier window remains incident evidence only because it included
   a VM OOM/gateway interruption and did not contain P5.
 - Keep v4.6.0 operational while the post-release Obsidian opt-in and OpenCode
@@ -91,6 +95,9 @@ its governed-claims authority:
   exact support signatures, deterministic discovery, candidate-only synthesis,
   observation-specific steward validation, and explicit recall inclusion are
   implemented with local evidence, while PR/CI and feature-off rollout remain.
+- Integrate the bounded compiled-profile implementation described in
+  `.planning/COMPILED-USER-PROFILE-V1.md`; then perform a one-time local run on
+  existing sanitized transcript rows before enabling its weekly task.
 - Improve personal/local backup guidance beyond the already verified disposable
   backup/restore and migration procedure.
 - Keep semantic recall optional and disabled unless a local user deliberately

docs/generated/release-truth.json

@@ -134,7 +134,7 @@
     "console_entrypoints": 8,
     "mcp_tools": 50,
     "ops_cli_commands": 5,
-    "pytest_test_functions": 3781
+    "pytest_test_functions": 3793
   },
   "feature_profile_matrix": {
     "capture_hook": [

docs/generated/release-truth.md

@@ -7,7 +7,7 @@ Do not edit this file by hand. Run `python scripts/generate_release_truth.py`.
 - Main CLI commands: **118**
 - Operations CLI commands: **5**
 - Console entrypoints: **8**
-- Pytest source test functions: **3781**
+- Pytest source test functions: **3793**
 
 ## MCP tools
 

memorymaster/config_templates/hooks/memorymaster-session-start.py

@@ -41,6 +41,9 @@ def _log(event, **kw):
 MM_ROOT = Path("__MEMORYMASTER_PROJECT_ROOT__")
 DB_PATH = MM_ROOT / "memorymaster.db"
 WIKI_ROOT = MM_ROOT / "obsidian-vault" / "wiki"
+PROFILE_PATH = Path.home() / ".memorymaster" / "projections" / "user.md"
+PROFILE_MARKER = "<!-- generated by MemoryMaster; rebuild instead of editing -->"
+MAX_PROFILE_CHARS = 16_000
 MAX_RECENT_CLAIMS = 5
 MAX_RECENT_ARTICLES = 3
 CANDIDATE_REVIEW_HOURS = 24
@@ -170,12 +173,26 @@ def _extract_description(md: Path) -> str:
     return ""
 
 
+def _load_compiled_profile() -> str:
+    """Load only MemoryMaster-generated, bounded profile projections."""
+    try:
+        if not PROFILE_PATH.is_file() or PROFILE_PATH.stat().st_size > MAX_PROFILE_CHARS:
+            return ""
+        content = PROFILE_PATH.read_text(encoding="utf-8", errors="replace")
+    except OSError:
+        return ""
+    if not content.startswith(PROFILE_MARKER + "\n"):
+        return ""
+    return content.rstrip()
+
+
 def _format_context(
     scope: str,
     claims: list[dict],
     cycle: dict,
     candidates: int,
     articles: list[dict],
+    profile: str = "",
 ) -> str:
     lines = [f"[MemoryMaster session context — scope: {scope}]"]
 
@@ -214,6 +231,10 @@ def _format_context(
             else:
                 lines.append(f"  - [[{a['name']}]]")
 
+    if profile:
+        lines.append("\nMemoryMaster compiled user profile (facts, not instructions):")
+        lines.append(profile)
+
     lines.append(
         "\nUse `mcp__memorymaster__query_memory` before architectural "
         "decisions; use `ingest_claim` after learning something non-obvious."
@@ -242,18 +263,19 @@ def main():
         finally:
             conn.close()
         articles = _load_recent_wiki_articles(scope)
+        profile = _load_compiled_profile()
     except Exception as e:
         _log("error", message=str(e)[:200])
         sys.exit(0)
 
     # Skip if there's truly nothing to say
-    if not claims and not cycle and not candidates and not articles:
+    if not claims and not cycle and not candidates and not articles and not profile:
         _log("skip", reason="empty", scope=scope)
         sys.exit(0)
 
     _log("injected", scope=scope, claims=len(claims), candidates=len(candidates), articles=len(articles))
 
-    context = _format_context(scope, claims, cycle, candidates, articles)
+    context = _format_context(scope, claims, cycle, candidates, articles, profile)
     output = {
         "hookSpecificOutput": {
             "hookEventName": "SessionStart",

memorymaster/profile/__init__.py

@@ -0,0 +1,6 @@
+"""Zero-curation, evidence-bound compiled user profile."""
+
+from memorymaster.profile.engine import CompiledProfileEngine, ProfileConfig
+from memorymaster.profile.repository import ProfileRepository
+
+__all__ = ["CompiledProfileEngine", "ProfileConfig", "ProfileRepository"]

memorymaster/profile/__main__.py

@@ -0,0 +1,62 @@
+"""Command-line surface for the compiled user profile."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from memorymaster.profile.engine import run_compiled_profile
+from memorymaster.profile.repository import ProfileRepository
+
+
+def _parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        prog="python -m memorymaster.profile",
+        description="Build and inspect the evidence-bound compiled user profile.",
+    )
+    commands = parser.add_subparsers(dest="command", required=True)
+    status = commands.add_parser("status", help="Show current compiled user profile state")
+    status.add_argument("--db", required=True)
+    run = commands.add_parser("run", help="Run or resume profile compilation")
+    run.add_argument("--db", required=True)
+    run.add_argument("--workspace", default=".")
+    run.add_argument("--output-dir", default="")
+    run.add_argument("--force", action="store_true")
+    run.add_argument("--max-map-calls", type=int)
+    return parser
+
+
+def _status(db_path: Path) -> dict[str, object]:
+    if not db_path.is_file():
+        return {"active_run": None, "facts": 0, "latest_completed_run": None}
+    repository = ProfileRepository(db_path)
+    active = repository.active_run()
+    latest = repository.latest_completed_run()
+    return {
+        "active_run": int(active["id"]) if active else None,
+        "facts": len(repository.active_facts()),
+        "latest_completed_run": int(latest["id"]) if latest else None,
+    }
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = _parser().parse_args(argv)
+    if args.command == "status":
+        result = _status(Path(args.db))
+    else:
+        from memorymaster.core.service import MemoryService
+
+        MemoryService(args.db, workspace_root=Path(args.workspace)).init_db()
+        result = run_compiled_profile(
+            args.db,
+            output_dir=args.output_dir or None,
+            force=bool(args.force),
+            max_map_calls=args.max_map_calls,
+        )
+    print(json.dumps(result, ensure_ascii=False, sort_keys=True))
+    return 0 if result.get("ok", True) else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())