7 天渐进式学习路径

import json, os

class FileMemory:
    # v2: 文件持久化 —— 进程重启后记忆仍在
    def __init__(self, path="memory.json"):
        self.path = path
        self.store = json.load(open(path)) if os.path.exists(path) else {}

    def remember(self, key, value):
        self.store[key] = value
        json.dump(self.store, open(self.path, "w"), ensure_ascii=False)

    def recall(self, key):
        return self.store.get(key, "（还没有相关记忆）")

mem = FileMemory()
mem.remember("user_name", "Kimi")
print(mem.recall("user_name"))  # → Kimi（重启后依然记得）

def build_context(system, history, retrieved, budget=8192):
    # 预算分配：system 20% / history 60% / retrieved 20%
    b_sys, b_his, b_ret = budget*0.2, budget*0.6, budget*0.2
    ctx = [trim(system, b_sys)]
    if tokens(history) > b_his:
        # 溢出 → 早期消息压缩为摘要
        head, tail = history[:-6], history[-6:]
        summary = llm_summarize(head)
        history = [sys_msg(f"前情提要：{summary}")] + tail
    ctx += trim(history, b_his)
    ctx += trim(retrieved, b_ret)
    return ctx

# RRF 混合检索：dense + sparse
def rrf_fuse(dense_hits, bm25_hits, k=60):
    scores = {}
    for rank, doc in enumerate(dense_hits):
        scores[doc.id] = scores.get(doc.id, 0) + 1/(k + rank)
    for rank, doc in enumerate(bm25_hits):
        scores[doc.id] = scores.get(doc.id, 0) + 1/(k + rank)
    return sorted(scores, key=scores.get, reverse=True)

# HNSW 索引关键参数
index = HNSW(M=16, efConstruction=200)  # M↑ 召回↑内存↑
results = index.search(q_vec, efSearch=64) # ef↑ 召回↑延迟↑

// 写入：用户-兴趣-领域
MERGE (u:User {name: $user})
MERGE (i:Interest {name: $interest})
MERGE (d:Domain {name: $domain})
MERGE (u)-[r:LIKES {since: datetime()}]->(i)
MERGE (i)-[:BELONGS_TO]->(d)

// 多跳推理：用户喜欢的领域下还有哪些兴趣？
MATCH (u:User {name: $user})-[:LIKES]->(i)-[:BELONGS_TO]->(d)
      <-[:BELONGS_TO]-(other:Interest)
WHERE NOT (u)-[:LIKES]->(other)
RETURN other.name ORDER BY count(d) DESC

def daily_consolidate(episodes):
    # 压缩漏斗：原始对话 → 事件摘要 → 信念提炼 → 画像更新
    L1 = [summarize(ep) for ep in episodes]      # 事件级
    L2 = extract_beliefs(L1)                    # 信念级
    profile = update_profile(L2)                # 画像级
    insights = llm(f"从以下记忆中生成3条洞察：\n{L2}")
    for ins in insights:
        ltm.write(ins, kind="insight",
                  confidence=score(ins))        # 低置信度不进上下文
    return profile, insights

class TeamMemory:
    def __init__(self):
        self.global_kb = {}   # 全局共享：世界知识
        self.team = {}        # 团队记忆：决策/规范/状态
        self.private = {}     # 个人记忆：{agent_id: {...}}

    def propose(self, agent, key, value):
        # 写入协议：提议 → 投票 → 提交
        proposal = {key: value, "by": agent, "votes": set()}
        return proposal

    def commit(self, proposal, voters, quorum=2):
        if len(proposal["votes"] & voters) >= quorum:
            self.team[next(iter(proposal))] = proposal
            self.broadcast("team_memory.updated", proposal)