M3a：息壤三级记忆域——归真触发提取管线/白名单明忆/重忆确认/正影版本链/删除墓碑/双通道召回注入，退役旧记忆-抽取-反思-大周天-演化管线（迁移 004，验收 7 用例）

2026-08-27 09:16:32 +08:00
1 parent 2a640a901b
commit 9c98c2f85e
@@ -0,0 +1,25 @@
+"""息壤（记忆域）：三级记忆、实录与召回（data-model.md §3）。"""
+
+from app.domains.xirang.service import (
+    confirm_memory,
+    correct_memory,
+    delete_memory,
+    deny_memory,
+    memorize_node,
+    on_canonized,
+    place_candidates,
+    recall_for_generation,
+    record_factual,
+)
+
+__all__ = [
+    "confirm_memory",
+    "correct_memory",
+    "delete_memory",
+    "deny_memory",
+    "memorize_node",
+    "on_canonized",
+    "place_candidates",
+    "recall_for_generation",
+    "record_factual",
+]
@@ -0,0 +1,504 @@
+"""息壤（记忆域）服务：提取管线、分级落位、确认流、正影、删除、检索与激活。
+
+铁律（memory-model.md §5）：只有归真结果（或用户手动入忆）能产生长期记忆候选——
+微痕不自动升级；未归真旁枝自动入忆率必须为 0（测试钉死）。
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import math
+import re
+from datetime import UTC, datetime
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core.errors import AppError
+from app.core.time import utc_now
+from app.db.session import AsyncSessionLocal
+from app.domains import one_qi
+from app.kernel import Kernel
+from app.models.mirror_tree import Canonization, TreeNode
+from app.models.xirang import FactualRecord, Memory, MemoryTrigger
+
+# 微痕默认半衰期（天）；命中强化 +0.08、上限 3.0（借鉴 Memory-Palace 活力模型）
+TRACE_HALF_LIFE_DAYS = 7.0
+VITALITY_HIT_BONUS = 0.08
+VITALITY_MAX = 3.0
+# 召回预算：条数与字符上限
+RECALL_LIMIT = 5
+RECALL_CHAR_BUDGET = 1200
+# 明忆白名单（PRD 三级记忆）：低敏客观类自动生效带撤销窗；心痕等主观类入候选箱
+EXPLICIT_AUTO_TYPES = ("persona", "instruction", "episodic")
+# deny 两次退出检索
+DENY_EXIT_THRESHOLD = 2
+
+deny_counts: dict[str, int] = {}
+
+
+# ── 归真触发：实录 + 提取 ──────────────────────────────────────────────
+
+
+async def on_canonized(canonization_id: str) -> None:
+    """归真后置（由 canonize 同步调用）：实录落库 + 调度异步提取。"""
+    await record_factual(canonization_id)
+    asyncio.get_running_loop().create_task(_extract_guarded(canonization_id))
+
+
+async def _extract_guarded(canonization_id: str) -> None:
+    try:
+        await _extract_after_canonized(canonization_id)
+    except Exception as exc:  # noqa: BLE001 — 后台任务异常必须可见
+        from app.core.logging import log_with_context
+
+        log_with_context(f"息壤提取任务失败：{exc}", level="ERROR")
+
+
+async def _extract_after_canonized(canonization_id: str) -> None:
+    async with AsyncSessionLocal() as session:
+        record = await session.get(Canonization, canonization_id)
+        if record is None:
+            return
+        path_ids = json.loads(record.path_nodes_json)
+        nodes = []
+        for nid in path_ids:
+            node = await session.get(TreeNode, nid)
+            if node is not None:
+                nodes.append(node)
+        transcript = "\n".join(
+            f"{'用户' if n.node_type == 'question' else '人物'}：{n.content}" for n in nodes
+        )
+        character_id = record.character_id
+    await _extract_candidates(
+        character_id=character_id, canonization_id=canonization_id, transcript=transcript
+    )
+
+
+async def record_factual(canonization_id: str) -> None:
+    """实录（同步，随归真事务后立即落库）。"""
+    async with AsyncSessionLocal() as session:
+        record = await session.get(Canonization, canonization_id)
+        if record is None:
+            return
+        path_nodes = []
+        for nid in json.loads(record.path_nodes_json):
+            node = await session.get(TreeNode, nid)
+            if node is not None:
+                path_nodes.append(node)
+        transcript = "\n".join(
+            f"{'用户' if n.node_type == 'question' else '人物'}：{n.content}" for n in path_nodes
+        )
+        summary = f"[{record.scope}] " + " / ".join(n.content[:40] for n in path_nodes[-2:])
+        session.add(
+            FactualRecord(
+                character_id=record.character_id,
+                canonization_id=canonization_id,
+                evidence_json=json.dumps({"path_nodes": [n.id for n in path_nodes]}, ensure_ascii=False),
+                summary=summary,
+            )
+        )
+        character_id = record.character_id
+        scope = record.scope
+        await session.commit()
+    _ = scope
+
+
+async def _extract_candidates(*, character_id: str, canonization_id: str, transcript: str) -> None:
+    """提取管线：LLM 抽取 JSON 候选；解析失败回落为确定性单条摘要候选（不阻塞闭环）。"""
+    from app.shenshi_client.types import ShenshiChatRequest, ShenshiMessage
+
+    kernel: Kernel | None = _kernel_ref()
+    raw = ""
+    if kernel is not None:
+        request = ShenshiChatRequest(
+            messages=[
+                ShenshiMessage(
+                    role="system",
+                    content=(
+                        "你是记忆抽取器。从对话实录中抽取长期记忆候选，"
+                        '输出 JSON 数组 [{"content","type","priority","confidence","sensitive"}]，'
+                        'type ∈ persona/episodic/instruction/relationship_view；宁缺毋滥，无值得记忆的内容输出 []。'
+                    ),
+                ),
+                ShenshiMessage(role="user", content=transcript[:8000]),
+            ],
+            run_id=f"extract_{canonization_id}",
+        )
+        parts: list[str] = []
+        try:
+            async for delta in kernel.shenshi.stream_chat(request):
+                parts.append(delta.delta)
+            raw = "".join(parts)
+        except Exception:  # noqa: BLE001 — 提取失败回落确定性路径
+            raw = ""
+
+    candidates = _parse_candidates(raw)
+    if not candidates:
+        candidates = [
+            {
+                "content": f"共同经历：{transcript[-200:]}",
+                "type": "episodic",
+                "priority": 50,
+                "confidence": 0.6,
+                "sensitive": False,
+            }
+        ]
+    await place_candidates(
+        character_id=character_id,
+        candidates=candidates,
+        source_refs=[canonization_id],
+        source="self_inferred",
+        realm="reality",
+    )
+
+
+def _parse_candidates(raw: str) -> list[dict]:
+    match = re.search(r"\[.*\]", raw, re.S)
+    if not match:
+        return []
+    try:
+        data = json.loads(match.group(0))
+    except json.JSONDecodeError:
+        return []
+    if not isinstance(data, list):
+        return []
+    out: list[dict] = []
+    for item in data:
+        if isinstance(item, dict) and item.get("content"):
+            out.append(item)
+    return out
+
+
+async def place_candidates(
+    *,
+    character_id: str,
+    candidates: list[dict],
+    source_refs: list[str],
+    source: str,
+    realm: str = "reality",
+) -> list[Memory]:
+    """分级落位 + 确定性去重（同类型同内容跳过）。"""
+    placed: list[Memory] = []
+    async with AsyncSessionLocal() as session:
+        existing = (
+            (
+                await session.execute(
+                    select(Memory).where(
+                        Memory.character_id == character_id, Memory.status.in_(("active", "pending"))
+                    )
+                )
+            )
+            .scalars()
+            .all()
+        )
+        existing_norms = {_norm(m.content) for m in existing}
+
+        for cand in candidates:
+            content = str(cand.get("content", "")).strip()
+            if not content or _norm(content) in existing_norms:
+                continue
+            sensitive = bool(cand.get("sensitive"))
+            mtype = str(cand.get("type", "episodic"))
+            if sensitive:
+                tier, status = "sensitive", "pending"  # 重忆必须确认（HJ-1201 管线侧）
+            elif mtype in EXPLICIT_AUTO_TYPES:
+                # 白名单明忆：低敏 persona/instruction/episodic 自动生效，撤销窗=可删除/可正影
+                tier, status = "explicit", "active"
+            else:
+                # 心痕（relationship_view，主观非事实）等入候选箱待审阅
+                tier, status = "explicit", "candidate"
+            memory = Memory(
+                character_id=character_id,
+                tier=tier,
+                type=mtype,
+                status=status,
+                source=source,
+                content=content,
+                source_refs_json=json.dumps(source_refs, ensure_ascii=False),
+                sensitivity="high" if sensitive else "low",
+                priority=int(cand.get("priority", 50)),
+                confidence=float(cand.get("confidence", 0.5)),
+                realm=realm,
+            )
+            session.add(memory)
+            await session.flush()
+            placed.append(memory)
+            for key in cand.get("keys", []) or []:
+                session.add(MemoryTrigger(memory_id=memory.id, kind="keyword", pattern=str(key)))
+        await session.commit()
+    for m in placed:
+        await one_qi.emit(
+            "memory.candidate_created",
+            character_id=character_id,
+            payload={"memory_id": m.id, "tier": m.tier, "status": m.status},
+            source="xirang",
+        )
+    return placed
+
+
+def _norm(text: str) -> str:
+    return re.sub(r"\s+", "", text).lower()
+
+
+def _kernel_ref() -> Kernel | None:
+    from app.main import app
+
+    return getattr(app.state, "kernel", None)
+
+
+# ── 手动入忆 ──────────────────────────────────────────────
+
+
+async def memorize_node(*, node_id: str, owner_scope: str, label: str) -> Memory:
+    """手动入忆：须显式标注 fact/insight/opinion/hypothesis；未归真内容只作 fact_candidate。"""
+    from app.models.character import Character
+
+    if label not in ("fact", "insight", "opinion", "hypothesis"):
+        raise AppError("invalid_request", "手动入忆必须标注 fact/insight/opinion/hypothesis。")
+    async with AsyncSessionLocal() as session:
+        node = await session.get(TreeNode, node_id)
+        if node is None:
+            raise AppError("not_found", "节点不存在。")
+        canonized_node_ids: set[str] = set()
+        canons = (
+            await session.execute(select(Canonization).where(Canonization.tree_id == node.tree_id))
+        ).scalars().all()
+        for canon in canons:
+            canonized_node_ids.update(json.loads(canon.path_nodes_json))
+        character = await session.get(Character, await _tree_character(session, node.tree_id))
+        if character is None or character.owner_scope != owner_scope:
+            raise AppError("not_found", "节点不存在。")
+        character_id = character.id
+
+        on_canon = node.id in canonized_node_ids
+        memory = Memory(
+            character_id=character.id,
+            tier="explicit",
+            type="episodic" if (on_canon and label == "fact") else "fact_candidate",
+            status="active",
+            source="user_explicit",
+            content=f"[{label}] {node.content}",
+            source_refs_json=json.dumps([node_id], ensure_ascii=False),
+            sensitivity="low",
+            realm=node.realm,
+        )
+        session.add(memory)
+        await session.flush()
+        memory_id = memory.id
+        await session.commit()
+    await one_qi.emit(
+        "memory.candidate_created",
+        character_id=character_id,
+        payload={"memory_id": memory_id, "manual": True, "label": label},
+        source="xirang",
+    )
+    async with AsyncSessionLocal() as session:
+        return await session.get(Memory, memory_id)  # type: ignore[return-value]
+
+
+async def _tree_character(session: AsyncSession, tree_id: str) -> str:
+    from app.models.mirror_tree import MirrorTree
+
+    tree = await session.get(MirrorTree, tree_id)
+    return tree.character_id if tree else ""
+
+
+# ── 确认流 / 正影 / 删除 ──────────────────────────────────────────────
+
+
+async def _owned_memory(memory_id: str, owner_scope: str, *, session: AsyncSession) -> Memory:
+    from app.models.character import Character
+
+    memory = await session.get(Memory, memory_id)
+    if memory is None:
+        raise AppError("not_found", "记忆不存在。")
+    character = await session.get(Character, memory.character_id)
+    if character is None or character.owner_scope != owner_scope:
+        raise AppError("not_found", "记忆不存在。")
+    return memory
+
+
+async def confirm_memory(memory_id: str, owner_scope: str) -> Memory:
+    async with AsyncSessionLocal() as session:
+        memory = await _owned_memory(memory_id, owner_scope, session=session)
+        if memory.status not in ("pending", "candidate"):
+            raise AppError("HJ-1103", "只有待确认（pending）或候选箱（candidate）的记忆可确认。", status_code=409)
+        memory.status = "active"
+        view = _view(memory)
+        await session.commit()
+    await one_qi.emit("memory.confirmed", character_id=view["character_id"], payload={"memory_id": memory_id}, source="xirang")
+    async with AsyncSessionLocal() as session:
+        return await session.get(Memory, memory_id)  # type: ignore[return-value]
+
+
+async def deny_memory(memory_id: str, owner_scope: str) -> dict:
+    async with AsyncSessionLocal() as session:
+        memory = await _owned_memory(memory_id, owner_scope, session=session)
+        memory.status = "tombstoned"
+        await session.commit()
+        character_id = memory.character_id
+    deny_counts[memory_id] = deny_counts.get(memory_id, 0) + 1
+    exited = deny_counts[memory_id] >= DENY_EXIT_THRESHOLD
+    await one_qi.emit(
+        "memory.denied", character_id=character_id, payload={"memory_id": memory_id, "exited": exited}, source="xirang"
+    )
+    return {"memory_id": memory_id, "exited_retrieval": True}
+
+
+async def correct_memory(memory_id: str, owner_scope: str, new_content: str) -> Memory:
+    """正影：新版本指向旧版本；旧版 superseded 停止普通召回，内容不改写。"""
+    async with AsyncSessionLocal() as session:
+        memory = await _owned_memory(memory_id, owner_scope, session=session)
+        if memory.status == "tombstoned":
+            raise AppError("HJ-1202", "已删除内容不可正影。", status_code=409)
+        new_version = Memory(
+            character_id=memory.character_id,
+            tier=memory.tier,
+            type=memory.type,
+            status="active" if memory.status in ("active", "pending") else memory.status,
+            source=memory.source,
+            content=new_content,
+            source_refs_json=memory.source_refs_json,
+            sensitivity=memory.sensitivity,
+            priority=memory.priority,
+            realm=memory.realm,
+            prev_version_id=memory.id,
+        )
+        session.add(new_version)
+        memory.status = "superseded"
+        await session.flush()
+        new_id = new_version.id
+        await session.commit()
+    await one_qi.emit(
+        "memory.superseded",
+        character_id=None,
+        payload={"memory_id": memory_id, "new_version_id": new_id},
+        source="xirang",
+    )
+    async with AsyncSessionLocal() as session:
+        return await session.get(Memory, new_id)  # type: ignore[return-value]
+
+
+async def delete_memory(memory_id: str, owner_scope: str) -> None:
+    """删除：墓碑化——内容不再出现在主存储查询/检索/上下文；墓碑不留原文。"""
+    async with AsyncSessionLocal() as session:
+        memory = await _owned_memory(memory_id, owner_scope, session=session)
+        character_id = memory.character_id
+        memory.status = "tombstoned"
+        memory.content = ""  # 墓碑不承载被删原文（data-model §10）
+        await session.commit()
+    await one_qi.emit(
+        "memory.deleted", character_id=character_id, payload={"memory_id": memory_id}, source="xirang"
+    )
+
+
+# ── 检索（双通道 + RRF 融合 + 预算）与激活 ──────────────────────────────────────────────
+
+
+async def recall_for_generation(
+    character_id: str, query_text: str, *, realm: str = "reality", limit: int = RECALL_LIMIT
+) -> list[Memory]:
+    """获准长期记忆召回：关键词通道 + 活力时近通道 → RRF 融合 → 预算裁剪。
+
+    通道 A：关键词/触发词命中（content LIKE 或 trigger pattern 命中）；
+    通道 B：活力值 × 时近衰减排序（标准档接入 pgvector 后升级为语义通道）。
+    过滤：active、界域匹配、非墓碑/被取代。
+    """
+    async with AsyncSessionLocal() as session:
+        actives = (
+            (
+                await session.execute(
+                    select(Memory).where(
+                        Memory.character_id == character_id,
+                        Memory.status == "active",
+                        Memory.realm.in_(("reality", realm)),
+                    )
+                )
+            )
+            .scalars()
+            .all()
+        )
+    if not actives:
+        return []
+
+    by_id = {m.id: m for m in actives}
+    now = utc_now()
+
+    def vitality_of(m: Memory) -> float:
+        age_days = max((now - _as_utc(m.created_at)).total_seconds() / 86400, 0)
+        return m.vitality * math.pow(2, -age_days / TRACE_HALF_LIFE_DAYS)
+
+    # 通道 A：关键词命中（含触发词）
+    keywords = {t.memory_id: t.pattern for t in await _triggers(set(by_id))}
+    channel_a: list[str] = []
+    text_norm = _norm(query_text)
+    for m in actives:
+        hit = _norm(m.content)[:24] in text_norm or any(
+            _norm(p) in text_norm for p in ([keywords.get(m.id)] if keywords.get(m.id) else [])
+        )
+        if hit or any(_norm(w) and _norm(w) in text_norm for w in _split_terms(m.content)):
+            channel_a.append(m.id)
+
+    # 通道 B：活力 × 时近
+    channel_b = sorted(by_id, key=lambda mid: vitality_of(by_id[mid]), reverse=True)
+
+    # RRF 融合
+    k = 60
+    scores: dict[str, float] = {}
+    for rank, mid in enumerate(channel_a):
+        scores[mid] = scores.get(mid, 0) + 1 / (k + rank + 1)
+    for rank, mid in enumerate(channel_b):
+        scores[mid] = scores.get(mid, 0) + 1 / (k + rank + 1)
+    ranked = sorted(scores, key=lambda mid: scores[mid], reverse=True)[:limit]
+
+    # 字符预算
+    picked: list[Memory] = []
+    used = 0
+    for mid in ranked:
+        m = by_id[mid]
+        if used + len(m.content) > RECALL_CHAR_BUDGET:
+            continue
+        picked.append(m)
+        used += len(m.content)
+
+    # 命中强化 + 召回统计
+    if picked:
+        async with AsyncSessionLocal() as session:
+            for m in picked:
+                row = await session.get(Memory, m.id)
+                row.vitality = min(row.vitality + VITALITY_HIT_BONUS, VITALITY_MAX)
+                row.recall_count += 1
+                row.last_recalled_at = utc_now()
+            await session.commit()
+    return picked
+
+
+async def _triggers(memory_ids: set[str]) -> list[MemoryTrigger]:
+    if not memory_ids:
+        return []
+    async with AsyncSessionLocal() as session:
+        result = await session.execute(
+            select(MemoryTrigger).where(MemoryTrigger.memory_id.in_(tuple(memory_ids)))
+        )
+        return list(result.scalars().all())
+
+
+def _split_terms(content: str) -> list[str]:
+    return [w for w in re.split(r"[，。！？、,.!?；;\s]+", content) if len(w) >= 2][:6]
+
+
+def _as_utc(dt: datetime) -> datetime:
+    return dt if dt.tzinfo else dt.replace(tzinfo=UTC)
+
+
+def _view(memory: Memory) -> dict:
+    return {
+        "id": memory.id,
+        "character_id": memory.character_id,
+        "tier": memory.tier,
+        "type": memory.type,
+        "status": memory.status,
+        "content": memory.content,
+    }