feat(memory): P1 长期记忆升级 —— 异步攒批 Consolidate + 软删 + importance/last_seen
把"逐轮盲写抽取"升级为 Mem0 式对账(方案见 memory_industry_analysis.md 落地节):
mcp-go:
- Profile 加 Importance(1~10, poignancy) + LastSeenAt(为 Generative Agents 读路径
Score=w1·Relevance+w2·Recency+w3·Importance 铺路)。
- Upsert 收 importance + 每次置 last_seen(印证);新增 Delete(软删,BaseModel.DeletedAt
已具备,失效不物删可审计)+ Touch;memory_upsert 透传 importance、新增 memory_delete 工具。
dispatcher:
- extractMemory → consolidateMemory:一次 LLM 调用同时做 抽取+对账,输出
[{op:ADD|UPDATE|DELETE|NOOP,key,value,importance}];ADD/UPDATE→upsert、DELETE→软删;
sanitizeOps 防幻删(DELETE 须命中已有)/夹 importance[1,10]/同key保末个/丢 NOOP。
- 攒批:每 3 轮(per-session 计数)才 consolidate 一次,省成本,对齐 ChatGPT 周期整理。
从根上解决 exact-key 盲写的记忆腐烂。
验证:parseOps/sanitizeOps/parseProfile 纯逻辑单测;store 集成测试(真 PG)覆盖
importance/last_seen 写入 + 软删(live 0 / 物理 1);dispatcher -race 全过。
(注:完整多轮 LLM consolidate 未做实跑,属构造性验证 + 沿用已证 pool.Chat 模式。)
P2 待做:读路径按 Score(Recency+Importance) 排序/衰减/截断 + 桌面端记忆面板。
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
@@ -7,74 +7,157 @@ import (
|
||||
"log"
|
||||
"strings"
|
||||
|
||||
"github.com/cloudwego/eino/schema"
|
||||
|
||||
"github.com/sundynix/sundynix-dispatcher/internal/llm"
|
||||
"github.com/sundynix/sundynix-shared/contract"
|
||||
)
|
||||
|
||||
// Pref 是从对话抽取出的一条长期偏好(key/value)。
|
||||
type Pref struct {
|
||||
Key string `json:"key"`
|
||||
Value string `json:"value"`
|
||||
// memOp 是 consolidate 的一条记忆操作(Mem0 式对账结果)。
|
||||
type memOp struct {
|
||||
Op string `json:"op"` // ADD / UPDATE / DELETE / NOOP
|
||||
Key string `json:"key"`
|
||||
Value string `json:"value"`
|
||||
Importance float64 `json:"importance"` // 1~10,poignancy(读路径权重)
|
||||
Reason string `json:"reason"`
|
||||
}
|
||||
|
||||
// extractMemory 写回阶段(异步、离开热路径):从本轮对话用 LLM 抽取用户长期偏好,
|
||||
// 与已有画像去重后经 memory_upsert 登记。模型/工具不可用或输入过短则跳过。
|
||||
func (o *Orchestrator) extractMemory(ctx context.Context, uid, input, answer string) {
|
||||
// consolidateMemory 写回阶段(异步、攒批):把近期对话与已有画像交给 LLM 一次对账,
|
||||
// 输出 ADD/UPDATE/DELETE/NOOP 操作并执行。ADD/UPDATE→upsert(带 importance)、DELETE→软删。
|
||||
// 一次 LLM 调用同时做"抽取+对账",从根上解决 exact-key 盲写导致的记忆腐烂。
|
||||
func (o *Orchestrator) consolidateMemory(ctx context.Context, uid string, hist []*schema.Message) {
|
||||
if uid == "" || o.tools == nil || o.pool == nil || !o.pool.Ready() {
|
||||
return
|
||||
}
|
||||
if len([]rune(strings.TrimSpace(input))) < 2 || len([]rune(strings.TrimSpace(answer))) < 20 {
|
||||
return // 太短,不值得抽取
|
||||
dialogue := renderDialogue(hist)
|
||||
if strings.TrimSpace(dialogue) == "" {
|
||||
return
|
||||
}
|
||||
existing := parseProfile(o.fetchMemory(ctx, uid, ""))
|
||||
|
||||
cctx, cancel := llmCtx(ctx)
|
||||
defer cancel()
|
||||
sys := "你从对话中提取【用户的长期稳定偏好或事实】(如称呼、语言、职业、专业领域、口味、常用工具、固定要求等)," +
|
||||
"忽略一次性的临时信息与你自己的话。"
|
||||
user := fmt.Sprintf("用户输入:%s\n助手回答:%s\n请抽取。只输出 JSON 数组 [{\"key\":\"偏好维度\",\"value\":\"值\"}],"+
|
||||
"没有可抽取的就输出 [],不要任何多余文字。", truncate(input, 800), truncate(answer, 1200))
|
||||
txt, err := o.pool.Chat(cctx, []llm.ChatMessage{{Role: "system", Content: sys}, {Role: "user", Content: user}})
|
||||
sys := "你是用户长期记忆管理器。对照【已有记忆】与【近期对话】,只针对用户长期稳定的偏好/事实做对账," +
|
||||
"忽略一次性临时信息。"
|
||||
txt, err := o.pool.Chat(cctx, []llm.ChatMessage{
|
||||
{Role: "system", Content: sys},
|
||||
{Role: "user", Content: consolidatePrompt(dialogue, existing)},
|
||||
})
|
||||
if err != nil {
|
||||
log.Printf("[eino] (writeback) 偏好抽取失败 user=%s: %v", uid, err)
|
||||
log.Printf("[eino] (writeback) 记忆对账失败 user=%s: %v", uid, err)
|
||||
return
|
||||
}
|
||||
fresh := filterNewPrefs(parsePrefs(txt), existing)
|
||||
for _, p := range fresh {
|
||||
o.upsertMemory(ctx, uid, p.Key, p.Value)
|
||||
ops := sanitizeOps(parseOps(txt), existing)
|
||||
var add, del int
|
||||
for _, op := range ops {
|
||||
switch op.Op {
|
||||
case "DELETE":
|
||||
o.deleteMemory(ctx, uid, op.Key)
|
||||
del++
|
||||
default: // ADD / UPDATE
|
||||
o.upsertMemory(ctx, uid, op.Key, op.Value, op.Importance)
|
||||
add++
|
||||
}
|
||||
}
|
||||
if len(fresh) > 0 {
|
||||
log.Printf("[eino] (writeback) 已登记 %d 条新偏好 user=%s", len(fresh), uid)
|
||||
if add+del > 0 {
|
||||
log.Printf("[eino] (writeback) 记忆对账 user=%s:写入/更新 %d,删除 %d", uid, add, del)
|
||||
}
|
||||
}
|
||||
|
||||
// upsertMemory 经 mcp-go memory_upsert 工具登记一条偏好。
|
||||
func (o *Orchestrator) upsertMemory(ctx context.Context, uid, key, value string) {
|
||||
cctx, cancel := context.WithTimeout(ctx, toolCallTimeout)
|
||||
defer cancel()
|
||||
if _, err := o.tools.CallTool(cctx, contract.ToolSubjectGo("memory_upsert"),
|
||||
&contract.ToolCall{Tool: "memory_upsert", Args: map[string]any{"user_id": uid, "key": key, "value": value}}); err != nil {
|
||||
log.Printf("[eino] memory_upsert 失败 %s=%s: %v", key, value, err)
|
||||
// consolidatePrompt 拼对账提示词:给出已有记忆 + 近期对话,要求输出操作列表。
|
||||
func consolidatePrompt(dialogue string, existing map[string]string) string {
|
||||
var ex strings.Builder
|
||||
if len(existing) == 0 {
|
||||
ex.WriteString("(暂无)")
|
||||
} else {
|
||||
for k, v := range existing {
|
||||
fmt.Fprintf(&ex, "- %s:%s\n", k, v)
|
||||
}
|
||||
}
|
||||
return fmt.Sprintf(`【已有记忆】
|
||||
%s
|
||||
【近期对话】
|
||||
%s
|
||||
|
||||
请对账并只输出 JSON 数组,每项 {"op":"ADD|UPDATE|DELETE|NOOP","key":"维度","value":"值","importance":1到10的整数,"reason":"简述"}:
|
||||
- ADD:对话中出现、已有记忆里没有的新长期偏好;
|
||||
- UPDATE:与已有某条同维度但值变了(**key 必须复用已有的那个**);
|
||||
- DELETE:已有记忆被对话明确推翻/过时(key 用已有的);
|
||||
- NOOP:无需变更的不要输出。
|
||||
importance 表示该偏好的重要程度(随口一提=低,强烈/反复强调=高)。只输出 JSON,无多余文字。`, strings.TrimRight(ex.String(), "\n"), dialogue)
|
||||
}
|
||||
|
||||
// parsePrefs 解析 LLM 抽取结果(容忍 ```json 围栏)为 []Pref,过滤空项。
|
||||
func parsePrefs(txt string) []Pref {
|
||||
var ps []Pref
|
||||
if json.Unmarshal([]byte(stripFence(txt)), &ps) != nil {
|
||||
// renderDialogue 把近期消息渲染为"用户/助手"对话文本(取末尾若干条,限长)。
|
||||
func renderDialogue(hist []*schema.Message) string {
|
||||
const maxMsgs = 8
|
||||
start := 0
|
||||
if len(hist) > maxMsgs {
|
||||
start = len(hist) - maxMsgs
|
||||
}
|
||||
var b strings.Builder
|
||||
for _, m := range hist[start:] {
|
||||
role := "用户"
|
||||
if m.Role == schema.Assistant {
|
||||
role = "助手"
|
||||
} else if m.Role == schema.System {
|
||||
continue
|
||||
}
|
||||
fmt.Fprintf(&b, "%s:%s\n", role, truncate(m.Content, 500))
|
||||
}
|
||||
return strings.TrimRight(b.String(), "\n")
|
||||
}
|
||||
|
||||
// parseOps 解析 LLM 对账结果(容忍 ```json 围栏)为 []memOp。
|
||||
func parseOps(txt string) []memOp {
|
||||
var ops []memOp
|
||||
if json.Unmarshal([]byte(stripFence(txt)), &ops) != nil {
|
||||
return nil
|
||||
}
|
||||
out := make([]Pref, 0, len(ps))
|
||||
for _, p := range ps {
|
||||
p.Key, p.Value = strings.TrimSpace(p.Key), strings.TrimSpace(p.Value)
|
||||
if p.Key != "" && p.Value != "" {
|
||||
out = append(out, p)
|
||||
return ops
|
||||
}
|
||||
|
||||
// sanitizeOps 校验/规整操作:规范 op 名;ADD/UPDATE 需 key+value;DELETE 需 key 且必须命中已有;
|
||||
// importance 夹到 [1,10];同 key 去重(保留末个);过滤 NOOP。
|
||||
func sanitizeOps(ops []memOp, existing map[string]string) []memOp {
|
||||
byKey := map[string]memOp{}
|
||||
order := []string{}
|
||||
for _, op := range ops {
|
||||
op.Op = strings.ToUpper(strings.TrimSpace(op.Op))
|
||||
op.Key = strings.TrimSpace(op.Key)
|
||||
op.Value = strings.TrimSpace(op.Value)
|
||||
if op.Key == "" {
|
||||
continue
|
||||
}
|
||||
switch op.Op {
|
||||
case "ADD", "UPDATE":
|
||||
if op.Value == "" {
|
||||
continue
|
||||
}
|
||||
if op.Importance < 1 {
|
||||
op.Importance = 1
|
||||
} else if op.Importance > 10 {
|
||||
op.Importance = 10
|
||||
}
|
||||
case "DELETE":
|
||||
if _, ok := existing[op.Key]; !ok { // 只删确实存在的,防误删/幻删
|
||||
continue
|
||||
}
|
||||
default: // NOOP 或未知 → 丢弃
|
||||
continue
|
||||
}
|
||||
if _, seen := byKey[op.Key]; !seen {
|
||||
order = append(order, op.Key)
|
||||
}
|
||||
byKey[op.Key] = op
|
||||
}
|
||||
out := make([]memOp, 0, len(order))
|
||||
for _, k := range order {
|
||||
out = append(out, byKey[k])
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// parseProfile 把 memory_get 渲染的画像("- 维度:值" 多行)解析回 map,供去重。
|
||||
// parseProfile 把 memory_get 渲染的画像("- 维度:值" 多行)解析回 map,供对账。
|
||||
func parseProfile(s string) map[string]string {
|
||||
m := map[string]string{}
|
||||
for _, line := range strings.Split(s, "\n") {
|
||||
@@ -95,19 +178,22 @@ func parseProfile(s string) map[string]string {
|
||||
return m
|
||||
}
|
||||
|
||||
// filterNewPrefs 保留新增或值有变化的偏好(同批同 key 去重;已有且相同则跳过)。
|
||||
func filterNewPrefs(extracted []Pref, existing map[string]string) []Pref {
|
||||
out := make([]Pref, 0, len(extracted))
|
||||
seen := map[string]bool{}
|
||||
for _, p := range extracted {
|
||||
if seen[p.Key] {
|
||||
continue
|
||||
}
|
||||
seen[p.Key] = true
|
||||
if cur, ok := existing[p.Key]; ok && cur == p.Value {
|
||||
continue
|
||||
}
|
||||
out = append(out, p)
|
||||
// upsertMemory 经 mcp-go memory_upsert 登记/更新一条偏好(带 importance)。
|
||||
func (o *Orchestrator) upsertMemory(ctx context.Context, uid, key, value string, importance float64) {
|
||||
cctx, cancel := context.WithTimeout(ctx, toolCallTimeout)
|
||||
defer cancel()
|
||||
if _, err := o.tools.CallTool(cctx, contract.ToolSubjectGo("memory_upsert"),
|
||||
&contract.ToolCall{Tool: "memory_upsert", Args: map[string]any{"user_id": uid, "key": key, "value": value, "importance": importance}}); err != nil {
|
||||
log.Printf("[eino] memory_upsert 失败 %s=%s: %v", key, value, err)
|
||||
}
|
||||
}
|
||||
|
||||
// deleteMemory 经 mcp-go memory_delete 软删一条偏好。
|
||||
func (o *Orchestrator) deleteMemory(ctx context.Context, uid, key string) {
|
||||
cctx, cancel := context.WithTimeout(ctx, toolCallTimeout)
|
||||
defer cancel()
|
||||
if _, err := o.tools.CallTool(cctx, contract.ToolSubjectGo("memory_delete"),
|
||||
&contract.ToolCall{Tool: "memory_delete", Args: map[string]any{"user_id": uid, "key": key}}); err != nil {
|
||||
log.Printf("[eino] memory_delete 失败 %s: %v", key, err)
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
@@ -2,19 +2,50 @@ package eino
|
||||
|
||||
import "testing"
|
||||
|
||||
func TestParsePrefs(t *testing.T) {
|
||||
got := parsePrefs("```json\n[{\"key\":\"称呼\",\"value\":\"Dexter\"},{\"key\":\"语言\",\"value\":\"中文\"},{\"key\":\"\",\"value\":\"空\"}]\n```")
|
||||
if len(got) != 2 {
|
||||
t.Fatalf("应解析出 2 条(过滤空 key),got %d: %v", len(got), got)
|
||||
func TestParseOps(t *testing.T) {
|
||||
got := parseOps("```json\n[{\"op\":\"ADD\",\"key\":\"称呼\",\"value\":\"Dexter\",\"importance\":7},{\"op\":\"DELETE\",\"key\":\"旧公司\"}]\n```")
|
||||
if len(got) != 2 || got[0].Op != "ADD" || got[0].Key != "称呼" || got[0].Importance != 7 || got[1].Op != "DELETE" {
|
||||
t.Fatalf("parseOps 解析错: %+v", got)
|
||||
}
|
||||
if got[0].Key != "称呼" || got[0].Value != "Dexter" {
|
||||
t.Errorf("解析错: %v", got[0])
|
||||
}
|
||||
if parsePrefs("不是 JSON") != nil {
|
||||
if parseOps("不是 JSON") != nil {
|
||||
t.Error("非 JSON 应返回 nil")
|
||||
}
|
||||
}
|
||||
|
||||
func TestSanitizeOps(t *testing.T) {
|
||||
existing := map[string]string{"语言": "中文", "旧公司": "A厂"}
|
||||
in := []memOp{
|
||||
{Op: "add", Key: "重要事", Value: "X", Importance: 99}, // 大小写规范 + importance 夹到 10
|
||||
{Op: "UPDATE", Key: "语言", Value: "英文", Importance: 0}, // importance 提到 1
|
||||
{Op: "ADD", Key: "空值", Value: ""}, // 无 value → 丢
|
||||
{Op: "DELETE", Key: "旧公司"}, // 命中已有 → 留
|
||||
{Op: "DELETE", Key: "不存在"}, // 未命中 → 丢(防幻删)
|
||||
{Op: "NOOP", Key: "语言"}, // NOOP → 丢(不覆盖已收的 UPDATE)
|
||||
{Op: "ADD", Key: "称呼", Value: "Dexter", Importance: 7},
|
||||
{Op: "ADD", Key: "称呼", Value: "Dex", Importance: 3}, // 同 key 重复 → 保留末个(值+importance)
|
||||
}
|
||||
out := sanitizeOps(in, existing)
|
||||
if len(out) != 4 {
|
||||
t.Fatalf("应剩 4 条(重要事/语言/删旧公司/称呼),得 %d: %+v", len(out), out)
|
||||
}
|
||||
byKey := map[string]memOp{}
|
||||
for _, o := range out {
|
||||
byKey[o.Key] = o
|
||||
}
|
||||
if byKey["重要事"].Importance != 10 {
|
||||
t.Errorf("重要事 importance 应夹到 10: %+v", byKey["重要事"])
|
||||
}
|
||||
if byKey["语言"].Op != "UPDATE" || byKey["语言"].Importance != 1 {
|
||||
t.Errorf("语言 应为 UPDATE 且 importance 提到 1: %+v", byKey["语言"])
|
||||
}
|
||||
if byKey["旧公司"].Op != "DELETE" {
|
||||
t.Errorf("旧公司 应保留 DELETE: %+v", byKey["旧公司"])
|
||||
}
|
||||
if byKey["称呼"].Value != "Dex" || byKey["称呼"].Importance != 3 {
|
||||
t.Errorf("称呼 应取末个 op(Dex/3): %+v", byKey["称呼"])
|
||||
}
|
||||
}
|
||||
|
||||
func TestParseProfile(t *testing.T) {
|
||||
m := parseProfile("- 称呼:Dexter\n- 语言: 中文\n\n- 职业:律师")
|
||||
if m["称呼"] != "Dexter" || m["语言"] != "中文" || m["职业"] != "律师" {
|
||||
@@ -24,20 +55,3 @@ func TestParseProfile(t *testing.T) {
|
||||
t.Error("空画像应得空 map")
|
||||
}
|
||||
}
|
||||
|
||||
func TestFilterNewPrefs(t *testing.T) {
|
||||
existing := map[string]string{"称呼": "Dexter", "语言": "中文"}
|
||||
in := []Pref{
|
||||
{"称呼", "Dexter"}, // 已有且相同 → 跳
|
||||
{"语言", "英文"}, // 已有但变了 → 留
|
||||
{"职业", "律师"}, // 新 → 留
|
||||
{"职业", "工程师"}, // 同批重复 key → 跳(保留首个)
|
||||
}
|
||||
got := filterNewPrefs(in, existing)
|
||||
if len(got) != 2 {
|
||||
t.Fatalf("应剩 2 条(语言变更 + 新职业),got %d: %v", len(got), got)
|
||||
}
|
||||
if got[0].Key != "语言" || got[0].Value != "英文" || got[1].Key != "职业" || got[1].Value != "律师" {
|
||||
t.Errorf("过滤结果不符: %v", got)
|
||||
}
|
||||
}
|
||||
|
||||
@@ -6,6 +6,7 @@ import (
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"log"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
"github.com/cloudwego/eino/schema"
|
||||
@@ -46,6 +47,9 @@ type Orchestrator struct {
|
||||
sink TokenSink
|
||||
tools ToolCaller
|
||||
exec ExecSink
|
||||
|
||||
turnMu sync.Mutex // 保护 turns(攒批计数,多任务 goroutine 共享)
|
||||
turns map[string]int // sessionID → 累计轮次,用于每 N 轮触发 consolidate
|
||||
}
|
||||
|
||||
// NewOrchestrator 持有依赖;图按任务的 DSL 在 Handle 内动态编译。
|
||||
@@ -169,8 +173,10 @@ func (o *Orchestrator) fetchHistory(ctx context.Context, sessionID string) []*sc
|
||||
return msgs
|
||||
}
|
||||
|
||||
// memorize 写回阶段:把本轮对话落进短期历史,并(TODO)抽取长期偏好记忆。
|
||||
// 异步执行,离开热路径。
|
||||
// consolidateEveryTurns 控制记忆对账的攒批节奏:每 N 轮 consolidate 一次(不逐轮,省成本)。
|
||||
const consolidateEveryTurns = 3
|
||||
|
||||
// memorize 写回阶段(异步、离热路径):落短期历史;每 N 轮做一次记忆对账(consolidate)。
|
||||
func (o *Orchestrator) memorize(t *contract.Task, answer string) {
|
||||
uid, _ := t.Meta[contract.MetaUserID].(string)
|
||||
sid, _ := t.Meta[contract.MetaSessionID].(string)
|
||||
@@ -179,9 +185,20 @@ func (o *Orchestrator) memorize(t *contract.Task, answer string) {
|
||||
o.appendHistory(sid, "assistant", answer)
|
||||
log.Printf("[eino] (writeback) task %s 已落会话历史 session=%s", t.ID, sid)
|
||||
}
|
||||
if uid != "" {
|
||||
// 从本轮对话抽取长期偏好 → 去重 → memory_upsert(离开热路径,已在 goroutine 内)。
|
||||
o.extractMemory(context.Background(), uid, dsl.Compile(t.Graph).Query, answer)
|
||||
if uid == "" || sid == "" || o.tools == nil {
|
||||
return
|
||||
}
|
||||
// 攒批:累计轮次,每 N 轮才把近期对话与已有画像交给 LLM 对账一次。
|
||||
o.turnMu.Lock()
|
||||
if o.turns == nil {
|
||||
o.turns = map[string]int{}
|
||||
}
|
||||
o.turns[sid]++
|
||||
n := o.turns[sid]
|
||||
o.turnMu.Unlock()
|
||||
if n%consolidateEveryTurns == 0 {
|
||||
ctx := context.Background()
|
||||
o.consolidateMemory(ctx, uid, o.fetchHistory(ctx, sid))
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user