dsh-lm-memory 0.0.0-stage → 0.18.0-rc.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +21 -0
- package/README.md +240 -2
- package/cordis.patch.yml +27 -0
- package/docs/banner-dark.svg +21 -0
- package/docs/banner.svg +21 -0
- package/docs/evidence/hardening/local-suite-summary.json +178 -0
- package/docs/evidence/hardening/rejected-precision-experiment.json +1315 -0
- package/docs/evidence/hardening/reliability-contracts.log +39 -0
- package/docs/evidence/hardening/semantic-after.json +1341 -0
- package/docs/evidence/hardening/semantic-before.json +1341 -0
- package/docs/evidence/hardening/source-hashes.json +12 -0
- package/docs/evidence/phase2/development-gate.json +53 -0
- package/docs/evidence/phase2/local-suite-summary.json +178 -0
- package/docs/evidence/phase2/longmemeval-full-summary.json +61 -0
- package/docs/evidence/phase2/semantic-regression.json +1341 -0
- package/docs/evidence/phase2/tail-proof.json +16 -0
- package/docs/evidence/recall-regression-2026-11-11.json +1304 -0
- package/docs/evidence/recall-suite-2026-11-11.json +126 -0
- package/docs/hardening-acceptance.md +50 -0
- package/docs/memory-protocol.md +89 -0
- package/docs/phase2-acceptance.md +51 -0
- package/docs/project-memory.md +28 -0
- package/docs/recall-contract-implementation.md +71 -0
- package/docs/task-recall-experiment.md +32 -0
- package/package.json +61 -4
- package/plugin/client.js +87 -0
- package/plugin/lib/confirmed-outcomes.mjs +37 -0
- package/plugin/lib/extract-helper.mjs +246 -0
- package/plugin/lib/index.js +506 -0
- package/plugin/lib/project-backends.mjs +56 -0
- package/plugin/lib/project-identity.mjs +30 -0
- package/plugin/lib/project-storage.mjs +62 -0
- package/plugin/lib/project-tool-catalog.json +339 -0
- package/plugin/lib/project-tools.mjs +58 -0
- package/plugin/lib/startup.mjs +50 -0
- package/plugin/lib/task-recall-worker.mjs +31 -0
- package/plugin/lib/task-recall.mjs +128 -0
- package/server/backup.mjs +16 -0
- package/server/candidates-e2e.mjs +221 -0
- package/server/conflict-e2e.mjs +134 -0
- package/server/conflict.mjs +153 -0
- package/server/current-md-e2e.mjs +73 -0
- package/server/dream-e2e.mjs +76 -0
- package/server/dream.mjs +43 -0
- package/server/e2e-test.mjs +44 -0
- package/server/extract-jobs-e2e.mjs +337 -0
- package/server/file-lock.mjs +16 -0
- package/server/graph-e2e.mjs +88 -0
- package/server/graph.mjs +125 -0
- package/server/guards.mjs +39 -0
- package/server/job-lease.mjs +10 -0
- package/server/lifecycle-e2e.mjs +141 -0
- package/server/memory-server.mjs +1224 -0
- package/server/passages.mjs +8 -0
- package/server/recall-eval.mjs +147 -0
- package/server/recall-utils.mjs +39 -0
- package/server/refresh-e2e.mjs +176 -0
- package/server/retrieval-plan.mjs +16 -0
- package/server/sleep-e2e.mjs +95 -0
- package/server/sleep.mjs +55 -0
- package/server/sparse-rank.mjs +14 -0
- package/server/viewer-e2e.mjs +89 -0
- package/server/viewer.mjs +725 -0
package/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 lm-memory contributors
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
package/README.md
CHANGED
|
@@ -1,3 +1,241 @@
|
|
|
1
|
-
|
|
1
|
+
<picture>
|
|
2
|
+
<source media="(prefers-color-scheme: dark)" srcset="docs/banner-dark.svg">
|
|
3
|
+
<img src="docs/banner.svg" alt="lm-memory — local-first long-term memory for agents">
|
|
4
|
+
</picture>
|
|
2
5
|
|
|
3
|
-
|
|
6
|
+
Agent 长期记忆:一个 MCP server、一个可选的 DSH 采集插件、一个本地管理页。单用户,SQLite + FTS5 + 本地双语嵌入(56MB)。存储和嵌入都在本机;提炼和摘要会调用你配置的远程 LLM,`allowRemote: false` 可以全部关掉。
|
|
7
|
+
|
|
8
|
+
这个项目被外部评审反复锤过十一轮(2026-11-17 起,5 分到 6.5 分再到逐轮收口),所以 README 里对可靠性、隐私、评测的描述都以故障注入测试和验收记录为准,不写做不到的事。
|
|
9
|
+
|
|
10
|
+
## 它解决什么问题
|
|
11
|
+
|
|
12
|
+
换个会话、换个宿主,AI 就失忆。lm-memory 给它们一块共享的、可检索、可纠正、会遗忘的记忆层。它最在意的不是功能多,而是三类常被忽略的故障:网络抖一下就永久漏记、续聊一次就覆盖掉新任务、错误记忆靠反复曝光给自己续命。这三个都有对应的机制和回归测试(见下文任务状态机、outbox、使用反馈续租)。
|
|
13
|
+
|
|
14
|
+
## 特性
|
|
15
|
+
|
|
16
|
+
| 能力 | 说明 |
|
|
17
|
+
|---|---|
|
|
18
|
+
| 统一提交入口 | MCP、采集卡、提炼候选三条写入路径走同一个事务函数 `commitMemoryTx`:凭据拦截 → 查重 → 事务 → 嵌入 → 索引 → 图谱 → 热流。提炼进程没有任何写库权限 |
|
|
19
|
+
| 持久 outbox | 内容变更和事件写在同一个 SQLite 事务里,事件写失败内容一起回滚。热流至少一次投递,每条事件带 `eid`,消费端去重 |
|
|
20
|
+
| 提炼任务状态机 | 派发、领取(owner + 10 分钟租约)、回执、结算全部落库。LLM 失败、截断输出、缺 key 都会保留任务重跑,超 5 次归档到 `extraction-jobs/failed/`。「有没有漏记」可以查表 |
|
|
21
|
+
| 使用反馈续租 | 召回命中只算曝光,`memory_feedback useful=true` 才续租。错误记忆没法靠反复出现在检索结果里给自己延寿 |
|
|
22
|
+
| 来源回查 | 自动条目保存脱敏后的输入窗口和消息序号,`memory_source` 可以查到这条记忆来自哪句话 |
|
|
23
|
+
| 同会话纠正 | 提炼输出可以带 `supersedes_id` 加逐字引文取代旧主张,引文对不上就不落库 |
|
|
24
|
+
| 检索 | FTS5 trigram 词法 + bge-small-zh/en 双语语义,按查询语言路由,RRF 融合。零命中会提示休眠馆里有别的记忆 |
|
|
25
|
+
| 遗忘 | 使用反馈 TTL + 按类型宽限期(待办 45 天、会话 60 天、事实和经验 120 天、偏好 240 天),超期转 dormant 降可见度,可唤醒,置顶豁免。这只是一套降可见度策略,没有证据说明它是正确的遗忘模型 |
|
|
26
|
+
| 批量写入 | `memory_batch_remember` 最多 50 条,逐条过统一提交防线 |
|
|
27
|
+
| 安全 | 凭据模式三处入口共用一套常量;采集落盘前脱敏;viewer 只接受本机严格同源请求;在线备份带 integrity 校验 |
|
|
28
|
+
| 记忆运维 | 冲突消解扫描(信号粗筛 + 模型复核,判定缓存化)、dream 去重、记忆图谱(similar/co_tag/ref 边)、变更热流向已开会话推送 |
|
|
29
|
+
|
|
30
|
+
## MCP 工具
|
|
31
|
+
|
|
32
|
+
| 工具 | 说明 |
|
|
33
|
+
|---|---|
|
|
34
|
+
| `memory_remember` | 写入。可显式传 `event_date`;凭据模式直接拒绝 |
|
|
35
|
+
| `memory_recall` | 词法+语义融合检索;`expand: true` 附图谱一跳扩展;零命中会提示休眠馆 |
|
|
36
|
+
| `memory_correct` | 修正,需 `expected_version`;新版本继承置顶和事件锚 |
|
|
37
|
+
| `memory_forget` | 默认逻辑删除;`hard: true` 按 id 物理删除任意状态的行(只删该行);休眠条目可删 |
|
|
38
|
+
| `memory_list` | 按种类列出最近记忆 |
|
|
39
|
+
| `memory_pin` | 置顶 / 取消置顶 / 唤醒休眠 |
|
|
40
|
+
| `memory_sleep` | 手动跑一轮休眠整理 + 冲突扫描 |
|
|
41
|
+
| `memory_feedback` | `{id, useful: true}` 更新续租统计。曝光不算使用 |
|
|
42
|
+
| `memory_source` | 查自动条目的原文来源(脱敏窗口 + 消息序号) |
|
|
43
|
+
| `memory_batch_remember` | 批量写入(≤50 条),逐条过完整防线 |
|
|
44
|
+
|
|
45
|
+
## 架构
|
|
46
|
+
|
|
47
|
+
```
|
|
48
|
+
你的 Agent(DSH / Claude Code / 其他 MCP 宿主)
|
|
49
|
+
│
|
|
50
|
+
├── MCP server(独立进程):10 个记忆工具 + 统一提交入口 + outbox + 任务账本
|
|
51
|
+
│
|
|
52
|
+
└── [仅 DSH] dsh-lm-capture 插件
|
|
53
|
+
会话结束 → 采集卡 + 提炼 job 文件 + 待派发通知(不直接 spawn)
|
|
54
|
+
↓
|
|
55
|
+
server dispatchJob:禁外发检查 → 登记 → 原子领取 → spawn helper
|
|
56
|
+
↓
|
|
57
|
+
helper:脱敏 → LLM 提炼/摘要 → 候选队列 + 回执
|
|
58
|
+
↓
|
|
59
|
+
server:候选消费 → 回执结算
|
|
60
|
+
周期任务:30 分钟任务回收 / 12h 休眠整理与补嵌
|
|
61
|
+
```
|
|
62
|
+
|
|
63
|
+
数据目录默认 `~/.dsh/memory/`(`LM_MEMORY_DIR` 或插件 `memoryDir` 可改):
|
|
64
|
+
|
|
65
|
+
| 文件 | 用途 |
|
|
66
|
+
|---|---|
|
|
67
|
+
| `memory.db` | SQLite + FTS5 + 向量 + outbox + 任务账本 |
|
|
68
|
+
| `current.md` | 注入层索引(置顶全文 + 一行指针;写失败时 EXISTS 重试自愈) |
|
|
69
|
+
| `log.jsonl` / `changes.jsonl` | 无内容审计流 / 带摘要的变更热流 |
|
|
70
|
+
| `pending-*.jsonl` | 采集卡、提炼候选、待派发任务队列(rename 快照崩溃恢复) |
|
|
71
|
+
| `extraction-jobs/` | 任务文件、`receipts/` 回执、`failed/` 归档 |
|
|
72
|
+
| `models/` | 嵌入模型缓存(首次下载约 56MB) |
|
|
73
|
+
|
|
74
|
+
## 任务怎么不丢
|
|
75
|
+
|
|
76
|
+
提炼任务的生命周期全部落库(`extract_jobs` 表),漏没漏记查表就知道:
|
|
77
|
+
|
|
78
|
+
```
|
|
79
|
+
登记(pending)
|
|
80
|
+
→ 原子领取(条件 UPDATE 抢 owner+租约,第二个 server 抢不到)
|
|
81
|
+
→ helper 执行(LLM 失败/截断/坏输出/缺 key → 保留任务,非零退出)
|
|
82
|
+
→ 候选消费(哪条失败哪条进重试队列,任务不标完成)
|
|
83
|
+
→ 回执结算(回执里全部候选 eid 都处置完才算 done;零候选也是明确终态)
|
|
84
|
+
→ 领取次数超 5 → failed 归档
|
|
85
|
+
```
|
|
86
|
+
|
|
87
|
+
几个设计决定:
|
|
88
|
+
|
|
89
|
+
- 候选 id 由 hash(sessionId · rev · kind · content) 确定性生成,rev 在任务创建时固定。同一个任务重跑不会换身份,重放不会把新摘要退回旧版;摘要更新还有版本序约束。
|
|
90
|
+
- 事件至少一次投递,靠 `eid` 在消费端去重;内容变更和事件同一个事务,不存在「库改了通知丢了」。
|
|
91
|
+
- 队列消费用 rename 快照,崩溃残留的 `.processing` 下次并回;写回失败就保留快照,从不删没落盘的数据。
|
|
92
|
+
- `allowRemote: false` 时派发入口在领取前就拦截,helper 自己也检查一遍;任务保留,恢复远程后自动重跑。
|
|
93
|
+
|
|
94
|
+
## 变更热流
|
|
95
|
+
|
|
96
|
+
写操作事件进 outbox,再投到 `changes.jsonl`,插件在每个模型 step 前按 per-session 字节 offset 增量读取,把变更摘要注进已开会话(一次最多 8 条)。候选队列的消费时机是 server 启动、召回前和周期巡检,不承诺会话运行中的秒级入库。
|
|
97
|
+
|
|
98
|
+
修改 `allowRemote` 后重启 DSH,使 MCP server 使用新配置。管理页是独立进程,只有能确认它由当前安装启动时才会自动替换;否则设置面板不会连接旧进程。请确认占用端口的是你启动的旧 lm-memory viewer,再手动停止并重启 DSH;若不是,请改用空闲端口。
|
|
99
|
+
|
|
100
|
+
## 安全
|
|
101
|
+
|
|
102
|
+
凭据防线是 `server/guards.mjs` 一个模块管三处入口(模式:`sk-…`、`api_key=…`、ghp/xox/AKIA 前缀):MCP 写入直接拒绝;采集在落盘之前脱敏;进远程 prompt 之前再过滤一遍,「别记」的消息整条剔除。评测分块摄入也走同样的双重过滤,有「分块拼不回凭据」的测试。
|
|
103
|
+
|
|
104
|
+
Viewer 监听回环地址,并检查请求的 Host。对携带 Origin 的请求,读接口允许本机来源和 `dsh-app://app`,在跨来源读取时回显 CORS;写接口只允许严格同源或 DSH Desktop 自定义协议来源。设置页先通过宿主接口确认当前 Viewer 地址与配置相符,再读取摘要。
|
|
105
|
+
|
|
106
|
+
边界要说清楚:凭据模式只认常见格式,不是机密检测器;逻辑删除保留审计行和历史热流摘要(删除事件本身不带原文);`hard: true` 按 id 删行,介质级擦除和备份清理不在此列。
|
|
107
|
+
|
|
108
|
+
## 评测
|
|
109
|
+
|
|
110
|
+
所有成绩都是 glm-5.3-flash 同模型作答和判分、bench 摄入而非生产自动提炼链路。跟任何厂商自报数字(Mem0 92.5%、Zep 94.7%)没有可比性。
|
|
111
|
+
|
|
112
|
+
### LoCoMo
|
|
113
|
+
|
|
114
|
+
| 协议 | 结果 |
|
|
115
|
+
|---|---|
|
|
116
|
+
| 前 3 段 55 题(官方类别映射勘误后重聚合) | 总体 40/55 = 0.727:multi-hop 0.67、temporal 0.93、open-domain 0.30、single-hop 0.87、adversarial 没测 |
|
|
117
|
+
| 全量 10 段 1986 题(2026-11-17) | 非陷阱 1540 题 73.77%(multi-hop 55.32、temporal 80.06、single-hop 81.21、open-domain 41.67);陷阱 446 题 82.06%;合计 75.63% |
|
|
118
|
+
|
|
119
|
+
历史小样当年用错了类别映射(官方映射见 [snap-research/locomo#29](https://github.com/snap-research/locomo/issues/29)),曾宣称的「multi-hop 0.93」「adversarial 0.87」都不成立。证据覆盖按三种口径算:任一命中 70.4%,真平均 62.5%,全部命中 55.6%——所以「检索已经解决」这个说法也撤回了。
|
|
120
|
+
|
|
121
|
+
### LongMemEval-S
|
|
122
|
+
|
|
123
|
+
| 协议 | 结果 |
|
|
124
|
+
|---|---|
|
|
125
|
+
| 旧 22 题 + 全量摘要时间线注入 | QA 0.55,gold session 覆盖@15 = 1.00(session 级命中不代表答案证据齐全) |
|
|
126
|
+
| 完整 S 500 题(2026-11-18,无时间线注入) | 62.0%:single-session-use 87.5%、multi-session 39.67%、temporal 59.06%、abstention 93.33% |
|
|
127
|
+
|
|
128
|
+
两行协议不同,差值不能叫改进。multi-session(跨会话合成)仍然是最弱的一项。
|
|
129
|
+
|
|
130
|
+
### 回归
|
|
131
|
+
|
|
132
|
+
`npm test` 跑 15 组套件:单元测试(含 reliability、来源回溯、批量防线)加 13 组 e2e(休眠、图谱、热流、任务状态机、并发领取、禁外发、隐私边界、viewer 边界)。检索实验的冻结验收记录在 [docs/phase2-acceptance.md](docs/phase2-acceptance.md)——四项消融没过预设门槛,相关实现默认关闭。
|
|
133
|
+
|
|
134
|
+
## 安装
|
|
135
|
+
|
|
136
|
+
Node.js ≥ 22.16(内置 `node:sqlite`,推荐 24+)。
|
|
137
|
+
|
|
138
|
+
### npm
|
|
139
|
+
|
|
140
|
+
```bash
|
|
141
|
+
npm install -g dsh-lm-memory
|
|
142
|
+
```
|
|
143
|
+
|
|
144
|
+
装好后把 server 挂进宿主(以 Claude Code 为例):
|
|
145
|
+
|
|
146
|
+
```bash
|
|
147
|
+
claude mcp add lm_memory -s user -- node "$(npm root -g)/dsh-lm-memory/server/memory-server.mjs"
|
|
148
|
+
```
|
|
149
|
+
|
|
150
|
+
### Claude Code 或任意 stdio MCP 宿主(git 方式)
|
|
151
|
+
|
|
152
|
+
```bash
|
|
153
|
+
git clone https://github.com/hrrrb408/lm-memory.git
|
|
154
|
+
claude mcp add lm_memory -s user -- node /path/to/lm-memory/server/memory-server.mjs
|
|
155
|
+
```
|
|
156
|
+
|
|
157
|
+
### DSH
|
|
158
|
+
|
|
159
|
+
测试候选版 `0.18.0-rc.2` 已发布到 GitHub Release,尚未发布到 npm。先核对 DSH 版本满足 `>=0.2.2-rc.3 <0.3.0`;在隔离环境打开 **插件 → 添加插件**,输入 `file:` 加下载到的 `dsh-lm-memory-0.18.0-rc.2.tgz` 绝对路径(Windows 例:`file:C:/LM-Test/dsh-lm-memory-0.18.0-rc.2.tgz`),按宿主提示重启。包只包含插件;安装依赖可能需要联网,且插件仍需要独立 Node.js ≥22.16。不要因此同意不明构建脚本或版本兼容豁免。
|
|
160
|
+
|
|
161
|
+
rc.2 的默认行为与 rc.1 相同。新增的按 Git 仓库隔离的项目记忆、用户确认结论/修复经验、事实有效日期,全部是显式开启的候选功能,默认关闭,说明与限制见 `docs/project-memory.md`。
|
|
162
|
+
|
|
163
|
+
以下是已发布稳定版示例。profile 的 `package.json` 加依赖(建议用 tag 固定版本):
|
|
164
|
+
|
|
165
|
+
```json
|
|
166
|
+
"dsh-lm-memory": "github:hrrrb408/lm-memory#v0.17.3"
|
|
167
|
+
```
|
|
168
|
+
|
|
169
|
+
`dsh.profile.bundles` 加 `"dsh-lm-memory"`,`pnpm install` 后重启 DSH。数据目录和安装位置无关,升级不动数据。
|
|
170
|
+
|
|
171
|
+
GUI 找不到可用 Node 时,将 `DSH_NODE` 设为符合上面版本要求的 Node 可执行文件绝对路径,再重启 DSH。新版启动逻辑在找不到可用 Node 时会跳过 MCP/viewer 启动并记录诊断,不会因此阻止宿主启动。
|
|
172
|
+
|
|
173
|
+
新版设置面板从 DSH 宿主读取 `viewer` 指定的本机端口(默认 `7640`),不会猜端口。若旧版 viewer 仍占用端口且无法确认进程归属,插件不会结束它;请手动停止自己启动的旧 viewer,再重启 DSH。
|
|
174
|
+
|
|
175
|
+
旧库(v0.16 表结构、hit_count 续租)启动时自动迁移:CHECK 约束重建,追加 `used_count/last_used_at` 列。旧曝光计数保留为统计,续租从零按使用反馈重新累计。
|
|
176
|
+
|
|
177
|
+
### 插件配置
|
|
178
|
+
|
|
179
|
+
```yaml
|
|
180
|
+
llm:
|
|
181
|
+
model: glm-5.3 # 默认跟随会话模型,显式指定即接管
|
|
182
|
+
# baseURL: "" # 换供应商才需要(OpenAI 兼容)
|
|
183
|
+
# apiKeyEnv: ZAI_CODING_CN_API_KEY
|
|
184
|
+
allowRemote: true # false = 系统级禁外发
|
|
185
|
+
memoryDir: "" # 自定义数据目录,队列/标记/任务路径跟随
|
|
186
|
+
refresh: true # false = 关闭变更热流注入
|
|
187
|
+
extract: true # false = 关闭自动提炼
|
|
188
|
+
# taskRecall: # 实验;不配置即关闭,需显式列出测试 Session ID
|
|
189
|
+
# enabled: true
|
|
190
|
+
# sessionIds: [session-替换为实际ID]
|
|
191
|
+
```
|
|
192
|
+
|
|
193
|
+
自动候选提示与 `refresh` 变更热流相互独立。启用方式、隐私与验收边界见[任务召回指针实验](docs/task-recall-experiment.md)。
|
|
194
|
+
|
|
195
|
+
key 从环境变量或 `~/.dsh/.credentials.yaml` 同名条目读取。没有 key 时自动提炼跳过但任务保留,重试 5 次后归档到 `extraction-jobs/failed/`,检索和手动写入不受影响。内置 provider 映射:zai-coding-cn、deepseek-official、deepseek-api-key、agnes。
|
|
196
|
+
|
|
197
|
+
### 使用反馈约定
|
|
198
|
+
|
|
199
|
+
休眠续租只认 `memory_feedback`,需要在宿主指令文件(比如 AGENTS.md)里加一条约定:
|
|
200
|
+
|
|
201
|
+
> 记忆被使用且确有帮助后,调用 `memory_feedback {id, useful: true}`;自动条目的重要主张先用 `memory_source` 回查原文。入库前先 recall 查重;已有正册档案的主题只存指针;当日流水账不入库。
|
|
202
|
+
|
|
203
|
+
### 评测复现
|
|
204
|
+
|
|
205
|
+
LoCoMo runner 的凭据走 `BENCH_BASE_URL` / `BENCH_API_KEY` / `BENCH_LLM`,默认不读个人凭据;`LM_BENCH_SPLIT=development|holdout|all` 控制对话范围。其他脚本在 `scripts/eval/`,源码哈希和原始判分输出随报告留档。
|
|
206
|
+
|
|
207
|
+
## 已知边界
|
|
208
|
+
|
|
209
|
+
- 遗忘这套是使用反馈 TTL,不是验证过的遗忘模型。反馈依赖宿主如实报告;置顶全文和启动索引的使用不计入。生产库还没有 dormant 样本,误休眠率和恢复率没有数据。
|
|
210
|
+
- 同会话的明确纠正能自动取代(要逐字引文),跨会话矛盾还是靠冲突扫描发现、人工复核,不做自动改写;语义改写后的同文防复活不保证。
|
|
211
|
+
- `event_date` 是稳定的观察锚,不等于考证过的事件日。旧数据回填自创建日,会有偏差;`memory_remember` 可以显式传。
|
|
212
|
+
- 没有独立 holdout 上的同条件竞品对照;judge 和答题是同一个模型;bench 摄入不等于生产链路。BM25 实验有个字符类缺陷待修。
|
|
213
|
+
- `memory_source` 只能回查到会话内消息,跨会话的实体和主张归一没做。自动提炼的内容没有经过语义验证,用之前自己判断。
|
|
214
|
+
- 图谱每次写操作全量重建(数百条毫秒级,数千条要换增量方案),对答案质量的收益没有消融数据,也不参与默认召回。
|
|
215
|
+
- 双语嵌入只是按查询语言路由,不是跨语言召回。bge-small 的窗口约 512 token,超长记忆(500 字以上)尾部语义会丢——两条 850 字、只有尾部城市不同的文本,向量完全一样。长内容拆开存。
|
|
216
|
+
- 单用户单机,无同步。删除的介质级擦除和备份清理不做。viewer 图谱页不支持触摸和键盘。
|
|
217
|
+
|
|
218
|
+
## 常见问题
|
|
219
|
+
|
|
220
|
+
**`ERR_PNPM_IGNORED_BUILDS`**:pnpm 10+ 默认不跑依赖构建脚本,onnxruntime-node 的二进制已在 tarball 里,功能正常,可忽略。想消除警告就在 `pnpm-workspace.yaml` 加 `onlyBuiltDependencies`。
|
|
221
|
+
|
|
222
|
+
**首次启动慢**:在下载双语嵌入模型(约 56MB)到 `models/`,正常。
|
|
223
|
+
|
|
224
|
+
**没有 LLM key**:自动提炼跳过但任务保留,检索、遗忘、viewer 都正常,手动 `memory_remember` 照常用。
|
|
225
|
+
|
|
226
|
+
**macOS 跑测试被 AMFI 拦**:DSH 托管的 node 会被拦 onnxruntime 原生库,本地测试换 Homebrew node;生产由宿主拉起,不受影响。
|
|
227
|
+
|
|
228
|
+
## 开发
|
|
229
|
+
|
|
230
|
+
```bash
|
|
231
|
+
npm test # 15 组套件
|
|
232
|
+
node scripts/eval/locomo-run.mjs 1 5 # LoCoMo 单段冒烟(需 key 和数据文件)
|
|
233
|
+
node scripts/dream-dryrun.mjs # 去重扫描(只读)
|
|
234
|
+
node scripts/conflict-dryrun.mjs # 冲突扫描(只读)
|
|
235
|
+
```
|
|
236
|
+
|
|
237
|
+
验收记录与未完成项:[docs/hardening-acceptance.md](docs/hardening-acceptance.md)、[docs/phase2-acceptance.md](docs/phase2-acceptance.md)。下一步待验收:真实 DSH 生命周期、跨会话事实归一、长期遗忘观察、同条件竞品对照。
|
|
238
|
+
|
|
239
|
+
## License
|
|
240
|
+
|
|
241
|
+
MIT
|
package/cordis.patch.yml
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
# dsh-lm-memory —— Living Memory 全家桶挂载声明(npm/GitHub 包形态)
|
|
2
|
+
# 默认数据目录由 DSH_HOME/系统用户目录定位;MCP 等待插件验证 Node 并写好稳定 shim。
|
|
3
|
+
# 平台无关(macOS / Windows / Linux)。注意:bundle 层只应用 insert: 块内的条目。
|
|
4
|
+
- insert:
|
|
5
|
+
- id: dsh-lm-memory
|
|
6
|
+
name: dsh-lm-memory
|
|
7
|
+
config:
|
|
8
|
+
minUserMessages: 2
|
|
9
|
+
memoryDir: !!js process.getBuiltinModule('node:path').join(process.env.DSH_HOME ?? process.getBuiltinModule('node:path').join(process.getBuiltinModule('node:os').homedir(), '.dsh'), 'memory')
|
|
10
|
+
viewer: 7640
|
|
11
|
+
# llm: 可选,提炼/摘要用的模型(不配则默认「跟随会话模型」;显式 model 一旦设置即接管)
|
|
12
|
+
# llm:
|
|
13
|
+
# model: glm-5.3
|
|
14
|
+
# baseURL: ""
|
|
15
|
+
# apiKeyEnv: ZAI_CODING_CN_API_KEY
|
|
16
|
+
# 已知可映射 provider:zai-coding-cn / deepseek-official / deepseek-api-key / agnes
|
|
17
|
+
- id: mcp-lm-memory
|
|
18
|
+
name: '@deepseek-ai/dsh-mcp-client'
|
|
19
|
+
inject: [lmMemoryStartup]
|
|
20
|
+
config:
|
|
21
|
+
serverName: lm_memory
|
|
22
|
+
transport: stdio
|
|
23
|
+
command: !!js ctx.lmMemoryStartup.node
|
|
24
|
+
args:
|
|
25
|
+
- !!js ctx.lmMemoryStartup.shimPath
|
|
26
|
+
env:
|
|
27
|
+
LM_MEMORY_DIR: !!js ctx.lmMemoryStartup.memoryDir
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 880 200" role="img" aria-label="lm-memory — local-first long-term memory for agents">
|
|
2
|
+
<style>
|
|
3
|
+
.bg { fill: #14171c; }
|
|
4
|
+
.ink { fill: #e8e4da; }
|
|
5
|
+
.sub { fill: #99a3ad; }
|
|
6
|
+
.accent { fill: #6f9bf2; }
|
|
7
|
+
.stroke { stroke: #e8e4da; }
|
|
8
|
+
.wire { stroke: #6f9bf2; }
|
|
9
|
+
</style>
|
|
10
|
+
<rect class="bg" x="0" y="0" width="880" height="200" rx="16"/>
|
|
11
|
+
<g fill="none" class="stroke" stroke-width="2">
|
|
12
|
+
<rect x="62" y="72" width="122" height="26" rx="9"/>
|
|
13
|
+
<rect x="76" y="44" width="108" height="26" rx="9"/>
|
|
14
|
+
<rect x="90" y="16" width="94" height="26" rx="9"/>
|
|
15
|
+
</g>
|
|
16
|
+
<line class="wire" x1="184" y1="29" x2="216" y2="29" stroke-width="2"/>
|
|
17
|
+
<circle class="accent" cx="222" cy="29" r="6"/>
|
|
18
|
+
<text class="ink" x="252" y="96" font-family="-apple-system,'Segoe UI',Roboto,'Helvetica Neue',Arial,sans-serif" font-size="58" font-weight="700" letter-spacing="-1.5">lm-memory</text>
|
|
19
|
+
<text class="sub" x="256" y="132" font-family="-apple-system,'Segoe UI',Roboto,'Helvetica Neue',Arial,sans-serif" font-size="21">local-first long-term memory for agents</text>
|
|
20
|
+
<text class="sub" x="256" y="166" font-family="ui-monospace,SFMono-Regular,Menlo,monospace" font-size="15">SQLite · FTS5 · bge embeddings · MCP · offline</text>
|
|
21
|
+
</svg>
|
package/docs/banner.svg
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 880 200" role="img" aria-label="lm-memory — local-first long-term memory for agents">
|
|
2
|
+
<style>
|
|
3
|
+
.bg { fill: #f7f3ea; }
|
|
4
|
+
.ink { fill: #232a32; }
|
|
5
|
+
.sub { fill: #68727c; }
|
|
6
|
+
.accent { fill: #2f6bd8; }
|
|
7
|
+
.stroke { stroke: #232a32; }
|
|
8
|
+
.wire { stroke: #2f6bd8; }
|
|
9
|
+
</style>
|
|
10
|
+
<rect class="bg" x="0" y="0" width="880" height="200" rx="16"/>
|
|
11
|
+
<g fill="none" class="stroke" stroke-width="2">
|
|
12
|
+
<rect x="62" y="72" width="122" height="26" rx="9"/>
|
|
13
|
+
<rect x="76" y="44" width="108" height="26" rx="9"/>
|
|
14
|
+
<rect x="90" y="16" width="94" height="26" rx="9"/>
|
|
15
|
+
</g>
|
|
16
|
+
<line class="wire" x1="184" y1="29" x2="216" y2="29" stroke-width="2"/>
|
|
17
|
+
<circle class="accent" cx="222" cy="29" r="6"/>
|
|
18
|
+
<text class="ink" x="252" y="96" font-family="-apple-system,'Segoe UI',Roboto,'Helvetica Neue',Arial,sans-serif" font-size="58" font-weight="700" letter-spacing="-1.5">lm-memory</text>
|
|
19
|
+
<text class="sub" x="256" y="132" font-family="-apple-system,'Segoe UI',Roboto,'Helvetica Neue',Arial,sans-serif" font-size="21">local-first long-term memory for agents</text>
|
|
20
|
+
<text class="sub" x="256" y="166" font-family="ui-monospace,SFMono-Regular,Menlo,monospace" font-size="15">SQLite · FTS5 · bge embeddings · MCP · offline</text>
|
|
21
|
+
</svg>
|
|
@@ -0,0 +1,178 @@
|
|
|
1
|
+
{
|
|
2
|
+
"node": "v26.12.0",
|
|
3
|
+
"created_at": "2026-11-17T08:54:03.195Z",
|
|
4
|
+
"results": [
|
|
5
|
+
{
|
|
6
|
+
"suite": "contracts",
|
|
7
|
+
"passed": true,
|
|
8
|
+
"exit_code": 0,
|
|
9
|
+
"duration_ms": 1686.28575,
|
|
10
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/contracts.log",
|
|
11
|
+
"evidence": [
|
|
12
|
+
"✔ capture retries failed dispatch and preserves message dates (15.650042ms)"
|
|
13
|
+
]
|
|
14
|
+
},
|
|
15
|
+
{
|
|
16
|
+
"suite": "sleep-e2e",
|
|
17
|
+
"passed": true,
|
|
18
|
+
"exit_code": 0,
|
|
19
|
+
"duration_ms": 317.32629199999997,
|
|
20
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/sleep-e2e.log",
|
|
21
|
+
"evidence": [
|
|
22
|
+
"✅ 审计含 sleep_pass/pin/wake",
|
|
23
|
+
"✅ schema 含 dormant 与新列",
|
|
24
|
+
"16 passed, 0 failed"
|
|
25
|
+
]
|
|
26
|
+
},
|
|
27
|
+
{
|
|
28
|
+
"suite": "dream-e2e",
|
|
29
|
+
"passed": true,
|
|
30
|
+
"exit_code": 0,
|
|
31
|
+
"duration_ms": 305.7172919999998,
|
|
32
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/dream-e2e.log",
|
|
33
|
+
"evidence": [
|
|
34
|
+
"✅ 高阈值下无误报 — total=0",
|
|
35
|
+
"✅ 跨 kind(fact×episode)照抓 — [[5,1,0.999],[5,2,0.9985],[2,1,0.995]]",
|
|
36
|
+
"8 passed, 0 failed"
|
|
37
|
+
]
|
|
38
|
+
},
|
|
39
|
+
{
|
|
40
|
+
"suite": "graph-e2e",
|
|
41
|
+
"passed": true,
|
|
42
|
+
"exit_code": 0,
|
|
43
|
+
"duration_ms": 335.8997919999997,
|
|
44
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/graph-e2e.log",
|
|
45
|
+
"evidence": [
|
|
46
|
+
"✅ expand 缺省仍返回数组老格式",
|
|
47
|
+
"✅ forget 后 A-B 边自动消失",
|
|
48
|
+
"14 passed, 0 failed"
|
|
49
|
+
]
|
|
50
|
+
},
|
|
51
|
+
{
|
|
52
|
+
"suite": "current-md-e2e",
|
|
53
|
+
"passed": true,
|
|
54
|
+
"exit_code": 0,
|
|
55
|
+
"duration_ms": 393.5648339999998,
|
|
56
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/current-md-e2e.log",
|
|
57
|
+
"evidence": [
|
|
58
|
+
"✅ 注入层 ≤ 25KB(60+5 条) — 7.6KB",
|
|
59
|
+
"✅ 注入层为索引形态(行数受限) — 80 行",
|
|
60
|
+
"12 passed, 0 failed"
|
|
61
|
+
]
|
|
62
|
+
},
|
|
63
|
+
{
|
|
64
|
+
"suite": "refresh-e2e",
|
|
65
|
+
"passed": true,
|
|
66
|
+
"exit_code": 0,
|
|
67
|
+
"duration_ms": 314.7228329999998,
|
|
68
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/refresh-e2e.log",
|
|
69
|
+
"evidence": [
|
|
70
|
+
" ✓ 派发通知写入配置目录",
|
|
71
|
+
" ✓ job 文件写入配置目录",
|
|
72
|
+
"=== 28 passed, 0 failed ==="
|
|
73
|
+
]
|
|
74
|
+
},
|
|
75
|
+
{
|
|
76
|
+
"suite": "viewer-e2e",
|
|
77
|
+
"passed": true,
|
|
78
|
+
"exit_code": 0,
|
|
79
|
+
"duration_ms": 4452.103834,
|
|
80
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/viewer-e2e.log",
|
|
81
|
+
"evidence": [
|
|
82
|
+
" ✓ 端点 pid 与实际端口占用一致",
|
|
83
|
+
" ✓ 新版在跑时再次 apply 不重启(pid 不变)",
|
|
84
|
+
"=== 5 passed, 0 failed ==="
|
|
85
|
+
]
|
|
86
|
+
},
|
|
87
|
+
{
|
|
88
|
+
"suite": "e2e-test",
|
|
89
|
+
"passed": true,
|
|
90
|
+
"exit_code": 0,
|
|
91
|
+
"duration_ms": 300.064042,
|
|
92
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/e2e-test.log",
|
|
93
|
+
"evidence": [
|
|
94
|
+
"[recall已删] 空 ✓"
|
|
95
|
+
]
|
|
96
|
+
},
|
|
97
|
+
{
|
|
98
|
+
"suite": "recall-eval",
|
|
99
|
+
"passed": true,
|
|
100
|
+
"exit_code": 0,
|
|
101
|
+
"duration_ms": 372.7857920000006,
|
|
102
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/recall-eval.log",
|
|
103
|
+
"evidence": [
|
|
104
|
+
"SUMMARY zhHard@5=0.50 zhEasy@5=0.75 en@5=0.80 traps=0/3"
|
|
105
|
+
]
|
|
106
|
+
},
|
|
107
|
+
{
|
|
108
|
+
"suite": "conflict-e2e",
|
|
109
|
+
"passed": true,
|
|
110
|
+
"exit_code": 0,
|
|
111
|
+
"duration_ms": 315.8230000000003,
|
|
112
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/conflict-e2e.log",
|
|
113
|
+
"evidence": [
|
|
114
|
+
"✅ parseReview 坏输出返回 null",
|
|
115
|
+
"✅ 纯关联引用(无取代语境)不出对",
|
|
116
|
+
"20 passed, 0 failed"
|
|
117
|
+
]
|
|
118
|
+
},
|
|
119
|
+
{
|
|
120
|
+
"suite": "lifecycle-e2e",
|
|
121
|
+
"passed": true,
|
|
122
|
+
"exit_code": 0,
|
|
123
|
+
"duration_ms": 322.93316700000105,
|
|
124
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/lifecycle-e2e.log",
|
|
125
|
+
"evidence": [
|
|
126
|
+
"✅ 留守员保持休眠",
|
|
127
|
+
"✅ 零命中提示休眠馆 — [](无相关记忆。不确定就不编造。) 另有 1 条低频记忆在休眠馆未参与本次检索——涉及较早的事实可用 scope:\"dormant\" 查询。",
|
|
128
|
+
"21 passed, 0 failed"
|
|
129
|
+
]
|
|
130
|
+
},
|
|
131
|
+
{
|
|
132
|
+
"suite": "candidates-e2e",
|
|
133
|
+
"passed": true,
|
|
134
|
+
"exit_code": 0,
|
|
135
|
+
"duration_ms": 1232.1287080000002,
|
|
136
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/candidates-e2e.log",
|
|
137
|
+
"evidence": [
|
|
138
|
+
"✅ 重跑后 s-legacy 摘要版本数仍为 2",
|
|
139
|
+
"43 passed, 0 failed",
|
|
140
|
+
"[memory] queue_merge_failed_proc_preserved {\"file\":\"pending-memories.jsonl\",\"error\":\"Error: EACCES: permission denied, open '/var/folders/l3/00y511gj4q55gj2z6zdqzrs0\"}"
|
|
141
|
+
]
|
|
142
|
+
},
|
|
143
|
+
{
|
|
144
|
+
"suite": "locomo-smoke",
|
|
145
|
+
"passed": true,
|
|
146
|
+
"exit_code": 0,
|
|
147
|
+
"duration_ms": 397.2574999999997,
|
|
148
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/locomo-smoke.log",
|
|
149
|
+
"evidence": [
|
|
150
|
+
"✅ 源码身份含 serverPath 与 gitHead 字段",
|
|
151
|
+
"✅ stderr 无 ReferenceError",
|
|
152
|
+
"8 passed, 0 failed"
|
|
153
|
+
]
|
|
154
|
+
},
|
|
155
|
+
{
|
|
156
|
+
"suite": "extract-jobs-e2e",
|
|
157
|
+
"passed": true,
|
|
158
|
+
"exit_code": 0,
|
|
159
|
+
"duration_ms": 27385.350417,
|
|
160
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/extract-jobs-e2e.log",
|
|
161
|
+
"evidence": [
|
|
162
|
+
"✅ 非法 Host 被拒(防 DNS rebinding) — status=403",
|
|
163
|
+
"46 passed, 0 failed",
|
|
164
|
+
"[memory] extract_job_failed_terminal {\"sessionId\":\"s-exhaust\",\"revision\":\"2026-11-17T04:03:00.000Z\",\"attempts\":6}"
|
|
165
|
+
]
|
|
166
|
+
},
|
|
167
|
+
{
|
|
168
|
+
"suite": "fixed-regression",
|
|
169
|
+
"passed": true,
|
|
170
|
+
"exit_code": 0,
|
|
171
|
+
"duration_ms": 419.6082500000048,
|
|
172
|
+
"log": "/var/folders/l3/00y511gj4q55gj2z6zdqzrs00000gn/T/lm-suite-4ZezdJ/fixed-regression.log",
|
|
173
|
+
"evidence": [
|
|
174
|
+
" \"annotation\": \"required_groups 中每组至少命中一条全文才算该组证据齐全;多组全部命中才算完整;relevant_ids 是允许相关的穷举标签。\""
|
|
175
|
+
]
|
|
176
|
+
}
|
|
177
|
+
]
|
|
178
|
+
}
|