memorymaster 3.25.0__tar.gz → 3.26.0__tar.gz

{memorymaster-3.25.0/memorymaster.egg-info → memorymaster-3.26.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memorymaster
-Version: 3.25.0
+Version: 3.26.0
 Summary: Production-grade memory reliability system for AI coding agents. Lifecycle-managed claims with citations, conflict detection, steward governance, and MCP integration.
 Author: wolverin0
 License: MIT
@@ -52,7 +52,7 @@ Lifecycle-managed claims with citations, conflict detection, steward governance,
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 [![Python 3.10+](https://img.shields.io/badge/python-3.10+-blue.svg)](https://www.python.org/downloads/)
-[![Tests](https://img.shields.io/badge/tests-2232-green.svg)]()
+[![Tests](https://img.shields.io/badge/tests-2514-green.svg)]()
 [![MCP Tools](https://img.shields.io/badge/MCP%20tools-24-purple.svg)]()
 [![CLI Commands](https://img.shields.io/badge/CLI%20commands-86-orange.svg)]()
 [![PyPI](https://img.shields.io/pypi/v/memorymaster.svg)](https://pypi.org/project/memorymaster/)

{memorymaster-3.25.0 → memorymaster-3.26.0}/README.md

@@ -6,7 +6,7 @@ Lifecycle-managed claims with citations, conflict detection, steward governance,
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 [![Python 3.10+](https://img.shields.io/badge/python-3.10+-blue.svg)](https://www.python.org/downloads/)
-[![Tests](https://img.shields.io/badge/tests-2232-green.svg)]()
+[![Tests](https://img.shields.io/badge/tests-2514-green.svg)]()
 [![MCP Tools](https://img.shields.io/badge/MCP%20tools-24-purple.svg)]()
 [![CLI Commands](https://img.shields.io/badge/CLI%20commands-86-orange.svg)]()
 [![PyPI](https://img.shields.io/pypi/v/memorymaster.svg)](https://pypi.org/project/memorymaster/)

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster/__init__.py

@@ -2,4 +2,4 @@
 
 __all__ = ["__version__"]
 
-__version__ = "3.25.0"
+__version__ = "3.26.0"

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster/contradiction_probe.py

@@ -77,9 +77,23 @@ def _canonical_pair(a_id: int, b_id: int) -> tuple[int, int]:
     return (a_id, b_id) if a_id <= b_id else (b_id, a_id)
 
 
-def _ensure_verdict_table(conn: sqlite3.Connection) -> None:
+_ensured_verdict_dbs: set[str] = set()
+
+
+def _ensure_verdict_table(conn: sqlite3.Connection, *, db_key: str | None = None) -> None:
+    """Create the verdict cache table if absent.
+
+    ``db_key`` lets per-claim callers skip the redundant ``CREATE TABLE IF NOT
+    EXISTS`` + ``commit`` on every claim in a steward cycle: once a given DB has
+    been ensured in this process we never re-issue the DDL. ``run_probe`` (one
+    DDL per whole run) passes no key and always ensures, preserving its semantics.
+    """
+    if db_key is not None and db_key in _ensured_verdict_dbs:
+        return
     conn.execute(_VERDICT_DDL)
     conn.commit()
+    if db_key is not None:
+        _ensured_verdict_dbs.add(db_key)
 
 
 def _cache_get(conn: sqlite3.Connection, a_id: int, b_id: int, model: str) -> dict | None:
@@ -164,13 +178,23 @@ def sample_candidate_pairs(
         return []
     embeddings = [provider.embed(_embed_text(c)) for c in usable]
     pairs: list[tuple[Claim, Claim, float]] = []
+    # The cosine sweep is O(n^2). Once we have already collected ``limit`` in-band
+    # pairs there is no point paying for the rest of the quadratic scan — the
+    # caller only ever consumes the first ``limit`` after the sort below, so break
+    # early instead of building (then discarding) thousands of extra pairs.
+    done = False
     for i in range(len(usable)):
+        if done:
+            break
         for j in range(i + 1, len(usable)):
             if _prefiltered(usable[i], usable[j]):
                 continue
             sim = cosine_similarity(embeddings[i], embeddings[j])
             if sim_low <= sim < sim_high:
                 pairs.append((usable[i], usable[j], round(sim, 4)))
+                if limit is not None and len(pairs) >= limit:
+                    done = True
+                    break
     pairs.sort(key=lambda p: -p[2])
     if limit is not None:
         pairs = pairs[:limit]
@@ -193,7 +217,11 @@ def _judge_llm(a: Claim, b: Claim) -> dict | None:
         if isinstance(item, dict) and "contradicts" in item:
             return {
                 "contradicts": bool(item.get("contradicts")),
-                "severity": (item.get("severity") or "low").strip().lower(),
+                # Floor empty/off-vocabulary severity to "medium" at the source
+                # (the JSON contract is advisory). Defaulting to "low" here would
+                # silently bury a real contradiction in the least-actionable tier
+                # and made the "medium" default downstream dead code.
+                "severity": _coerce_severity(item.get("severity")),
                 "reason": (item.get("reason") or "").strip(),
                 "cached": False,
             }
@@ -206,6 +234,36 @@ def _model_key() -> str:
     return f"{provider}:{model}"
 
 
+_VALID_SEVERITIES = {"low", "medium", "high"}
+
+
+def _coerce_severity(value: Any) -> str:
+    """Normalize a verdict severity to a canonical bucket.
+
+    The LLM judge and verdict cache can both yield an empty or off-vocabulary
+    severity (the JSON contract is advisory, not enforced). When that happens we
+    floor to ``"medium"`` rather than ``"low"`` so a contradiction surfaced by
+    the probe is never silently downgraded to the least-actionable tier.
+    """
+    sev = str(value or "").strip().lower()
+    return sev if sev in _VALID_SEVERITIES else "medium"
+
+
+def _safe_judge_reason(reason: Any) -> str | None:
+    """Return the raw judge reason unless it leaks sensitive content.
+
+    Both ``probe_for_claim`` and ``run_probe(apply=True)`` persist/surface the
+    LLM judge reason. Routing both through this guard ensures a reason that the
+    sensitivity filter flags is dropped (returns ``None``) on EVERY path — the
+    events table must never receive an unredacted secret.
+    """
+    text = str(reason or "")
+    _, leaks = redact_text(text)
+    if leaks:
+        return None
+    return text
+
+
 # ---------------------------------------------------------------------------
 # Wilson confidence interval
 # ---------------------------------------------------------------------------
@@ -303,15 +361,19 @@ def run_probe(
                 if verdict["contradicts"]:
                     stats["contradictions"] += 1
                     loser, winner = (a, b) if a.confidence <= b.confidence else (b, a)
+                    # Same sensitivity guard the per-claim path uses: never let an
+                    # unredacted judge reason reach the events table on apply.
+                    safe_reason = _safe_judge_reason(verdict.get("reason"))
                     stats["found"].append({
                         "claim_a_id": a.id, "claim_b_id": b.id, "similarity": sim,
-                        "severity": verdict["severity"], "reason": verdict["reason"],
+                        "severity": verdict["severity"], "reason": verdict.get("reason", ""),
                         "flag_candidate_id": loser.id,
                     })
                     if apply:
+                        detail = f" ({safe_reason})" if safe_reason else ""
                         transition_claim(
                             service.store, loser.id, "conflicted",
-                            reason=f"contradiction_probe: contradicts claim {winner.id} ({verdict['reason']})",
+                            reason=f"contradiction_probe: contradicts claim {winner.id}{detail}",
                             event_type="transition",
                         )
                         stats["flagged_conflicted"] += 1
@@ -415,7 +477,7 @@ def probe_for_claim(
 
     conn = sqlite3.connect(str(db_path))
     try:
-        _ensure_verdict_table(conn)
+        _ensure_verdict_table(conn, db_key=str(db_path))
         model = _model_key()
         for peer, sim in candidates:
             metrics["pairs_checked"] += 1
@@ -439,14 +501,13 @@ def probe_for_claim(
                 _cache_put(conn, claim.id, peer.id, model, verdict)
             if not verdict.get("contradicts"):
                 continue
-            verdict_reason = verdict.get("reason", "")
-            _, leaks = redact_text(verdict_reason)
-            if leaks:
+            verdict_reason = _safe_judge_reason(verdict.get("reason"))
+            if verdict_reason is None:
                 continue  # drop rather than surface a sensitive judge-reason
             metrics["contradictions"] += 1
             reasons.append({
                 "code": "contradiction_probe.semantic_pair",
-                "severity": verdict.get("severity", "medium"),
+                "severity": _coerce_severity(verdict.get("severity")),
                 "detail": f"Semantic contradiction with claim {peer.id}: {verdict_reason}",
                 "evidence": {
                     "peer_claim_id": int(peer.id),

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster/query_cache.py

@@ -116,7 +116,14 @@ def read(db_path: str, cache_key: str) -> list[dict] | None:
             "SELECT result_json, generation FROM query_cache WHERE cache_key = ?",
             (cache_key,),
         ).fetchone()
-        if row is None or int(row["generation"]) != gen:
+        if row is None:
+            return None
+        if int(row["generation"]) != gen:
+            try:
+                conn.execute("DELETE FROM query_cache WHERE cache_key = ?", (cache_key,))
+                conn.commit()
+            except sqlite3.Error:
+                pass
             return None
         return json.loads(row["result_json"])
     except (sqlite3.Error, json.JSONDecodeError, ValueError) as exc:
@@ -126,6 +133,23 @@ def read(db_path: str, cache_key: str) -> list[dict] | None:
         conn.close()
 
 
+def evict_stale(db_path: str) -> None:
+    """Delete query cache rows older than the current corpus generation."""
+    try:
+        conn = _connect(db_path)
+    except sqlite3.Error as exc:
+        logger.warning("query_cache.evict_stale connect failed: %s", exc)
+        return
+    try:
+        gen = current_generation(conn)
+        conn.execute("DELETE FROM query_cache WHERE generation < ?", (gen,))
+        conn.commit()
+    except (sqlite3.Error, ValueError) as exc:
+        logger.warning("query_cache.evict_stale failed: %s", exc)
+    finally:
+        conn.close()
+
+
 def write(db_path: str, cache_key: str, stub_rows: list[dict], generation: int) -> None:
     """Store result stubs tagged with ``generation``. Best-effort.
 

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster/rule_miner.py

@@ -48,6 +48,11 @@ _CORRECTION_KEYWORDS = (
     "wrong", "not what", "that's not", "thats not", "revert", "undo",
     "should have", "shouldn't", "why did you", "no need", "stop ",
 )
+_CORRECTION_FTS_MATCH = (
+    '"no" OR "don" OR "do" OR "dont" OR "instead" OR "actually" OR '
+    '"wrong" OR "not" OR "revert" OR "undo" OR "should" OR "shouldn" OR '
+    '"why" OR "stop"'
+)
 
 _MINER_STATE_DDL = """
 CREATE TABLE IF NOT EXISTS miner_state (
@@ -132,6 +137,18 @@ def _candidate_batch(
     params: list[Any] = [since_id]
     params.extend(f"%{kw}%" for kw in _CORRECTION_KEYWORDS)
     params.append(batch_size)
+    try:
+        return conn.execute(
+            f"""SELECT v.id, v.session_id, v.content, v.scope FROM verbatim_fts f
+                JOIN verbatim_memories v ON v.id = f.rowid
+                WHERE verbatim_fts MATCH ?
+                  AND v.role = 'user' AND v.id > ? AND ({like_clause})
+                ORDER BY v.id ASC LIMIT ?""",
+            [_CORRECTION_FTS_MATCH, *params],
+        ).fetchall()
+    except sqlite3.Error as exc:
+        logger.debug("rule_miner: verbatim_fts pre-filter failed, using LIKE scan: %s", exc)
+
     return conn.execute(
         f"""SELECT id, session_id, content, scope FROM verbatim_memories
             WHERE role = 'user' AND id > ? AND ({like_clause})

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster/verbatim_cleanup.py

@@ -132,10 +132,19 @@ def cleanup(
 
         if dedup:
             # Identify ids to drop: for each (session_id, content) group, keep
-            # the smallest id (oldest), drop the rest.
+            # the smallest id (oldest), drop the rest. NOT EXISTS against a
+            # correlated "is there an older twin?" probe avoids the O(n^2)
+            # full-table anti-join that `id NOT IN (SELECT MIN(id) ...)` forces
+            # SQLite into on the cold CLI path. Results are identical: a row is
+            # dropped iff another row in the same (session_id, content) group has
+            # a strictly smaller id.
             rows = conn.execute(
-                """SELECT id FROM verbatim_memories WHERE id NOT IN (
-                       SELECT MIN(id) FROM verbatim_memories GROUP BY session_id, content
+                """SELECT id FROM verbatim_memories AS v
+                   WHERE EXISTS (
+                       SELECT 1 FROM verbatim_memories AS older
+                       WHERE older.session_id IS v.session_id
+                         AND older.content = v.content
+                         AND older.id < v.id
                    )"""
             ).fetchall()
             ids = [int(r[0]) for r in rows]

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster/verbatim_store.py

@@ -22,11 +22,11 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any
 
+# Credential detection delegated to the canonical filter in memorymaster.security.
+from memorymaster.security import redact_text as _redact_text
+
 logger = logging.getLogger(__name__)
 
-# Credential detection delegated to the canonical filter in
-# memorymaster.security — single source of truth.
-from memorymaster.security import redact_text as _redact_text
 
 
 def _contains_sensitive(text: str) -> bool:
@@ -76,6 +76,33 @@ def store_verbatim(
     timestamp: str | None = None,
 ) -> int | None:
     """Store a verbatim conversation turn. Returns row ID or None if filtered."""
+    # closing() guarantees the connection (and its WAL write lock) is released
+    # even if an INSERT/commit raises (e.g. "database is locked" under
+    # concurrent MCP/Stop-hook writers) - this is the hottest write path.
+    with closing(_connect(db_path)) as conn:
+        row_id = _store_verbatim_conn(
+            conn,
+            session_id,
+            role,
+            content,
+            scope,
+            source_agent,
+            timestamp,
+        )
+        conn.commit()
+        return row_id
+
+
+def _store_verbatim_conn(
+    conn: sqlite3.Connection,
+    session_id: str,
+    role: str,
+    content: str,
+    scope: str = "project",
+    source_agent: str = "",
+    timestamp: str | None = None,
+) -> int | None:
+    """Store one turn using an existing connection without committing."""
     if not content or len(content) < 20:
         return None
     if _row_has_sensitive_field(role or "", source_agent or "", content):
@@ -83,37 +110,32 @@ def store_verbatim(
 
     now = timestamp or datetime.now(timezone.utc).isoformat()
 
-    # closing() guarantees the connection (and its WAL write lock) is released
-    # even if an INSERT/commit raises (e.g. "database is locked" under
-    # concurrent MCP/Stop-hook writers) — this is the hottest write path.
-    with closing(_connect(db_path)) as conn:
-        # Dedup by exact content within the same session. Uses idx_verbatim_session
-        # so the lookup is O(rows-in-session), not table-scan. The previous FTS5-based
-        # dedup query passed a sha256 hex prefix to MATCH which never resolves —
-        # the FTS5 index stores the content text, not its hash. Result: 9M+ rows
-        # accumulated for orchestrator sessions because every Stop event re-inserted
-        # every message. Fixed 2026-05-03; see mm-0c43.
-        existing = conn.execute(
-            "SELECT id FROM verbatim_memories WHERE session_id = ? AND content = ? LIMIT 1",
-            (session_id, content),
-        ).fetchone()
-        if existing:
-            return None
-
-        cur = conn.execute(
-            """INSERT INTO verbatim_memories (session_id, role, content, scope, timestamp, source_agent)
-               VALUES (?, ?, ?, ?, ?, ?)""",
-            (session_id, role, content, scope, now, source_agent),
-        )
-        row_id = cur.lastrowid
+    # Dedup by exact content within the same session. Uses idx_verbatim_session
+    # so the lookup is O(rows-in-session), not table-scan. The previous FTS5-based
+    # dedup query passed a sha256 hex prefix to MATCH which never resolves -
+    # the FTS5 index stores the content text, not its hash. Result: 9M+ rows
+    # accumulated for orchestrator sessions because every Stop event re-inserted
+    # every message. Fixed 2026-05-03; see mm-0c43.
+    existing = conn.execute(
+        "SELECT id FROM verbatim_memories WHERE session_id = ? AND content = ? LIMIT 1",
+        (session_id, content),
+    ).fetchone()
+    if existing:
+        return None
 
-        # Update FTS
-        conn.execute(
-            "INSERT INTO verbatim_fts(rowid, content) VALUES (?, ?)",
-            (row_id, content),
-        )
-        conn.commit()
-        return row_id
+    cur = conn.execute(
+        """INSERT INTO verbatim_memories (session_id, role, content, scope, timestamp, source_agent)
+           VALUES (?, ?, ?, ?, ?, ?)""",
+        (session_id, role, content, scope, now, source_agent),
+    )
+    row_id = cur.lastrowid
+
+    # Update FTS
+    conn.execute(
+        "INSERT INTO verbatim_fts(rowid, content) VALUES (?, ?)",
+        (row_id, content),
+    )
+    return row_id
 
 
 def _extract_role_content(entry: dict) -> tuple[str, str]:
@@ -164,29 +186,31 @@ def store_transcript(
     stats = {"stored": 0, "skipped": 0}
     session_id = path.stem  # Use filename as session ID
 
-    for line in path.read_text(encoding="utf-8", errors="replace").splitlines():
-        if not line.strip():
-            continue
-        try:
-            entry = json.loads(line)
-        except json.JSONDecodeError:
-            continue
-        if not isinstance(entry, dict):
-            continue
-
-        role, content = _extract_role_content(entry)
-        if role not in ("user", "assistant"):
-            stats["skipped"] += 1
-            continue
-        if not content or len(content) < 20:
-            stats["skipped"] += 1
-            continue
-
-        row_id = store_verbatim(db_path, session_id, role, content, scope, source_agent)
-        if row_id:
-            stats["stored"] += 1
-        else:
-            stats["skipped"] += 1
+    with closing(_connect(db_path)) as conn:
+        for line in path.read_text(encoding="utf-8", errors="replace").splitlines():
+            if not line.strip():
+                continue
+            try:
+                entry = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+            if not isinstance(entry, dict):
+                continue
+
+            role, content = _extract_role_content(entry)
+            if role not in ("user", "assistant"):
+                stats["skipped"] += 1
+                continue
+            if not content or len(content) < 20:
+                stats["skipped"] += 1
+                continue
+
+            row_id = _store_verbatim_conn(conn, session_id, role, content, scope, source_agent)
+            if row_id:
+                stats["stored"] += 1
+            else:
+                stats["skipped"] += 1
+        conn.commit()
 
     return stats
 

{memorymaster-3.25.0 → memorymaster-3.26.0/memorymaster.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memorymaster
-Version: 3.25.0
+Version: 3.26.0
 Summary: Production-grade memory reliability system for AI coding agents. Lifecycle-managed claims with citations, conflict detection, steward governance, and MCP integration.
 Author: wolverin0
 License: MIT
@@ -52,7 +52,7 @@ Lifecycle-managed claims with citations, conflict detection, steward governance,
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 [![Python 3.10+](https://img.shields.io/badge/python-3.10+-blue.svg)](https://www.python.org/downloads/)
-[![Tests](https://img.shields.io/badge/tests-2232-green.svg)]()
+[![Tests](https://img.shields.io/badge/tests-2514-green.svg)]()
 [![MCP Tools](https://img.shields.io/badge/MCP%20tools-24-purple.svg)]()
 [![CLI Commands](https://img.shields.io/badge/CLI%20commands-86-orange.svg)]()
 [![PyPI](https://img.shields.io/pypi/v/memorymaster.svg)](https://pypi.org/project/memorymaster/)

{memorymaster-3.25.0 → memorymaster-3.26.0}/memorymaster.egg-info/SOURCES.txt

@@ -218,6 +218,7 @@ tests/test_auto_ingest_hook_schema.py
 tests/test_auto_resolver.py
 tests/test_auto_validate.py
 tests/test_backend_parity.py
+tests/test_bench_answer_prompt.py
 tests/test_bm25_per_field.py
 tests/test_calibration.py
 tests/test_calibration_priors_applied.py
@@ -229,6 +230,7 @@ tests/test_classify_hook_f1.py
 tests/test_classify_hook_latency.py
 tests/test_claude_to_turns.py
 tests/test_cli_dry_run.py
+tests/test_cli_handlers_extra.py
 tests/test_cli_json_flag.py
 tests/test_cli_ready.py
 tests/test_cli_review_queue.py
@@ -245,6 +247,7 @@ tests/test_confusion_matrix_eval.py
 tests/test_connection_retry.py
 tests/test_connectors.py
 tests/test_context_hook.py
+tests/test_context_hook_extra.py
 tests/test_context_optimizer.py
 tests/test_context_optimizer_provider.py
 tests/test_contradiction_probe.py
@@ -258,6 +261,7 @@ tests/test_dashboard_review_queue.py
 tests/test_daydream_ingest.py
 tests/test_db_merge_confidence_conflict.py
 tests/test_db_merge_coverage_v2.py
+tests/test_db_merge_extra.py
 tests/test_decay_coverage.py
 tests/test_decay_respects_pinned.py
 tests/test_dedup.py
@@ -266,6 +270,7 @@ tests/test_dedup_conflict_disambiguation.py
 tests/test_delta_sync.py
 tests/test_deterministic_predicates.py
 tests/test_dream_bridge_coverage_v2.py
+tests/test_dream_bridge_extra.py
 tests/test_dream_bridge_sensitivity.py
 tests/test_embeddings_coverage.py
 tests/test_entity_extractor.py
@@ -275,6 +280,7 @@ tests/test_entity_graph_export.py
 tests/test_entity_new_kinds.py
 tests/test_entity_regex_v3.py
 tests/test_entity_registry.py
+tests/test_entity_registry_extra.py
 tests/test_eval_harness.py
 tests/test_events_schema.py
 tests/test_extract_llm_ollama.py
@@ -296,6 +302,7 @@ tests/test_lifecycle_supersede_invariant.py
 tests/test_llm_budget.py
 tests/test_llm_fallback.py
 tests/test_llm_provider_claude_cli.py
+tests/test_llm_provider_extra.py
 tests/test_llm_provider_key_rotation.py
 tests/test_llm_steward_coverage.py
 tests/test_llm_steward_key_rotation.py
@@ -310,6 +317,7 @@ tests/test_meta_decisions.py
 tests/test_metrics_exporter.py
 tests/test_migrations.py
 tests/test_observability.py
+tests/test_observability_extra.py
 tests/test_obsidian_mind_patterns.py
 tests/test_operator.py
 tests/test_operator_queue.py
@@ -332,6 +340,7 @@ tests/test_recall_tokenizer.py
 tests/test_recall_vector_fallback.py
 tests/test_reliability_hardening.py
 tests/test_resolvers_concurrent_supersede.py
+tests/test_retrieval_extra.py
 tests/test_retrieval_profile.py
 tests/test_retrieval_profiles.py
 tests/test_retrieval_rrf_tiebreaker.py
@@ -364,6 +373,7 @@ tests/test_steward_daydream_hook.py
 tests/test_steward_features.py
 tests/test_steward_features_v3.py
 tests/test_steward_resolution_parity.py
+tests/test_storage_extra.py
 tests/test_storage_parity.py
 tests/test_store_factory.py
 tests/test_tenant_isolation.py
@@ -375,6 +385,7 @@ tests/test_v313_run_cycle_dedupe.py
 tests/test_v390_e2e.py
 tests/test_v391_strict_warnings.py
 tests/test_vault_exporter.py
+tests/test_vault_linter_extra.py
 tests/test_vault_linter_orphan.py
 tests/test_vector_search.py
 tests/test_verbatim_cleanup.py
@@ -387,6 +398,7 @@ tests/test_webhook_hmac.py
 tests/test_whatsapp_importer.py
 tests/test_wiki_autopromote.py
 tests/test_wiki_binding.py
+tests/test_wiki_engine_extra.py
 tests/test_wiki_engine_idempotency.py
 tests/test_wiki_explored_and_contradictions.py
 tests/test_wiki_freshness.py

{memorymaster-3.25.0 → memorymaster-3.26.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "memorymaster"
-version = "3.25.0"
+version = "3.26.0"
 description = "Production-grade memory reliability system for AI coding agents. Lifecycle-managed claims with citations, conflict detection, steward governance, and MCP integration."
 license = {text = "MIT"}
 authors = [{name = "wolverin0"}]

{memorymaster-3.25.0 → memorymaster-3.26.0}/tests/bench_longmemeval.py

@@ -696,8 +696,9 @@ def call_claude_cli_judge(
 def answer_question(question: str, contexts: list[str], judge: JudgeClient) -> str:
     prompt = "\n\n".join(
         [
-            "Answer the question using only the retrieved conversation context.",
-            "If the context does not contain the answer, say I don't know.",
+            "Answer the question using the retrieved conversation context. "
+            "Answer with the most specific relevant fact from the retrieved context; "
+            "only say I don't know if the answer is genuinely absent.",
             "Keep the answer concise.",
             "Retrieved context:",
             "\n\n---\n\n".join(contexts),

memorymaster-3.26.0/tests/test_bench_answer_prompt.py

@@ -0,0 +1,75 @@
+"""Guards the LongMemEval answerer prompt against over-refusal regressions.
+
+WHY: The bench's `answer_question` prompt drives the QA-accuracy metric. An
+earlier wording ("If the context does not contain the answer, say I don't
+know.") biased the answerer toward refusing whenever the fact was not stated
+verbatim, depressing QA accuracy on questions whose answer WAS present but
+phrased differently. The contract for this prompt is now: extract the most
+specific relevant fact from the retrieved context, and only abstain when the
+answer is genuinely absent. These tests anchor on that requirement (intent),
+not on the exact prose, so they stay meaningful if the wording is polished
+further — what must hold is that the prompt instructs specificity and scopes
+the abstention to genuine absence rather than mere non-verbatim phrasing.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+_SPEC = importlib.util.spec_from_file_location(
+    "bench_longmemeval",
+    Path(__file__).resolve().parent / "bench_longmemeval.py",
+)
+assert _SPEC and _SPEC.loader
+bench = importlib.util.module_from_spec(_SPEC)
+# Register before exec so the module's frozen @dataclass decorators can resolve
+# their own __module__ via sys.modules during class creation.
+sys.modules[_SPEC.name] = bench
+_SPEC.loader.exec_module(bench)
+
+
+class _PromptCapturingJudge:
+    """Captures the prompt instead of calling any provider."""
+
+    def __init__(self) -> None:
+        self.prompt: str | None = None
+
+    def complete(self, prompt, *, max_tokens, temperature=0.0):
+        self.prompt = prompt
+        return bench.LLMResponse(text="captured", model="test", provider="test")
+
+
+def _build_prompt() -> str:
+    judge = _PromptCapturingJudge()
+    bench.answer_question("What city did the user move to?", ["ctx-a", "ctx-b"], judge)
+    assert judge.prompt is not None
+    return judge.prompt
+
+
+def test_prompt_requests_specific_fact() -> None:
+    # Intent: the answerer must be told to surface the most specific relevant
+    # fact, not just "answer the question" generically.
+    prompt = _build_prompt().lower()
+    assert "most specific" in prompt
+    assert "relevant fact" in prompt
+
+
+def test_abstention_is_scoped_to_genuine_absence() -> None:
+    # Intent: abstention ("I don't know") must be gated on the answer being
+    # genuinely absent — NOT the blanket "if the context does not contain the
+    # answer" wording that triggered over-refusal on non-verbatim matches.
+    prompt = _build_prompt().lower()
+    assert "genuinely absent" in prompt
+    assert "if the context does not contain the answer" not in prompt
+
+
+def test_prompt_still_includes_question_and_contexts() -> None:
+    # Guard the structural contract so the softening did not drop inputs.
+    judge = _PromptCapturingJudge()
+    bench.answer_question("Q?", ["alpha-ctx", "beta-ctx"], judge)
+    assert judge.prompt is not None
+    assert "Q?" in judge.prompt
+    assert "alpha-ctx" in judge.prompt
+    assert "beta-ctx" in judge.prompt