@restregular/dsh-memory 0.0.0-stage → 0.1.0
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 +69 -2
- package/cordis.patch.yml +10 -0
- package/lib/inject/apply.js +82 -0
- package/lib/inject/inject.js +127 -0
- package/lib/provider/apply.js +108 -0
- package/lib/provider/embed-server.js +296 -0
- package/lib/provider/embed.js +105 -0
- package/lib/provider/index.js +612 -0
- package/lib/provider/rerank-cache.js +64 -0
- package/lib/provider/rerank.js +169 -0
- package/lib/provider/scoring.js +145 -0
- package/lib/provider/secrets.js +57 -0
- package/lib/provider/store.js +485 -0
- package/lib/tools/apply.js +106 -0
- package/lib/tools/index.js +224 -0
- package/package.json +49 -4
package/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 restregular
|
|
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 +1,70 @@
|
|
|
1
|
-
#
|
|
1
|
+
# @restregular/dsh-memory
|
|
2
2
|
|
|
3
|
-
|
|
3
|
+
DeepSeek Harness 的本地长期记忆插件(核心包):JSONL 存储 + 本地向量检索 + 四个模型工具 + 自动召回注入。
|
|
4
|
+
|
|
5
|
+
## 安装
|
|
6
|
+
|
|
7
|
+
装进任意 DSH profile(例如桌面端 `~/.dsh/profiles/desktop`,Web 端 `~/.dsh/profiles/web`):
|
|
8
|
+
|
|
9
|
+
```bash
|
|
10
|
+
pnpm add @restregular/dsh-memory
|
|
11
|
+
```
|
|
12
|
+
|
|
13
|
+
本包同时提供三个 loader entry,装好后自动生效(bundle patch 已随包分发):
|
|
14
|
+
|
|
15
|
+
| entry id | 作用 |
|
|
16
|
+
| --- | --- |
|
|
17
|
+
| `dsh-memory-local` | 提供 `memory` 服务:存储、检索、向量回填 |
|
|
18
|
+
| `dsh-tool-memory` | 注册 `memory_save` / `memory_search` / `memory_list` / `memory_forget` 四个工具 |
|
|
19
|
+
| `dsh-memory-inject` | 在 `agent/pre-step` 注入召回块(可选,可关) |
|
|
20
|
+
|
|
21
|
+
> 管理面板是独立包:**@restregular/dsh-memory-panel**(设置页里查看/编辑/删除记忆,并手动启停向量服务)。
|
|
22
|
+
|
|
23
|
+
## 配置
|
|
24
|
+
|
|
25
|
+
patch 里给 `dsh-memory-local` 传配置即可(默认值通常够用):
|
|
26
|
+
|
|
27
|
+
```yaml
|
|
28
|
+
- id: dsh-memory-local
|
|
29
|
+
config:
|
|
30
|
+
storeDir: '~/.dsh/memory-store' # 默认 <home>/.dsh/memory-store
|
|
31
|
+
scopes: { global: true, project: true }
|
|
32
|
+
retrieval:
|
|
33
|
+
topK: 5
|
|
34
|
+
vector:
|
|
35
|
+
enabled: true
|
|
36
|
+
endpoint: 'http://127.0.0.1:8078/v1/embeddings'
|
|
37
|
+
model: bge-m3
|
|
38
|
+
dims: 1024
|
|
39
|
+
timeoutMs: 300 # 交互检索预算(要紧)
|
|
40
|
+
writeTimeoutMs: 10000 # 保存时嵌整条记录
|
|
41
|
+
backfillTimeoutMs: 30000 # 批量回填
|
|
42
|
+
backfillMaxChars: 8000
|
|
43
|
+
llm: { enabled: false } # LLM 重排(可选)
|
|
44
|
+
limits: { maxRecordsPerScope: 2000, maxRecordChars: 4000 }
|
|
45
|
+
secrets: { enabled: true, mode: refuse } # 拒绝写入疑似凭据
|
|
46
|
+
```
|
|
47
|
+
|
|
48
|
+
### 本地向量服务(可选)
|
|
49
|
+
|
|
50
|
+
不配向量服务时,检索自动降级为纯词法(IDF + CJK bigram),功能不受影响。要启用语义检索,指向任意 OpenAI 兼容的 `/v1/embeddings` 服务(推荐 `llama-server` + BGE-M3):
|
|
51
|
+
|
|
52
|
+
```yaml
|
|
53
|
+
vector:
|
|
54
|
+
spawn: # 让插件托管 llama-server 进程(opt-in)
|
|
55
|
+
enabled: true
|
|
56
|
+
exe: '/path/to/llama-server.exe'
|
|
57
|
+
model: '/path/to/bge-m3-Q4_K_M.gguf'
|
|
58
|
+
cwd: '/path/to/llamacpp' # 该目录须含 llama-server-impl.dll
|
|
59
|
+
logFile: '/path/to/llama-server.log'
|
|
60
|
+
threads: 6
|
|
61
|
+
```
|
|
62
|
+
|
|
63
|
+
`spawn.enabled: true` 时:应用启动后台拉起、应用退出自动关闭;已在跑的实例(有 PID 文件)会被接管,没有 PID 文件的外部实例不接管。也可在面板里手动开关。
|
|
64
|
+
|
|
65
|
+
## 说明
|
|
66
|
+
|
|
67
|
+
- 存储为每作用域一个 JSONL + `.vec` 向量边车(含模型/维度身份头,不匹配自动重建)。
|
|
68
|
+
- 记忆分 `global` 与 `project:<cwd>` 两类作用域。
|
|
69
|
+
- 只在插件自己的生命周期内工作:任何异常都被吞掉并记日志,绝不让 DSH 启动失败。
|
|
70
|
+
- License: MIT
|
package/cordis.patch.yml
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
1
|
+
# @restregular/dsh-memory bundle patch: local memory provider + model-facing tools + recall injection.
|
|
2
|
+
# 三个 entry 共用本包,但本包不声明 dsh.client —— 声明了客户端 bundle 的包只允许一个活跃
|
|
3
|
+
# entry(dsh-client-modules: "resolves from multiple active Loader sources")。
|
|
4
|
+
- insert:
|
|
5
|
+
- id: dsh-memory-local
|
|
6
|
+
name: '@restregular/dsh-memory'
|
|
7
|
+
- id: dsh-tool-memory
|
|
8
|
+
name: '@restregular/dsh-memory/tools'
|
|
9
|
+
- id: dsh-memory-inject
|
|
10
|
+
name: '@restregular/dsh-memory/inject'
|
|
@@ -0,0 +1,82 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cordis wiring for dsh-memory-inject. Loaded only inside DSH: hooks the
|
|
3
|
+
* documented agent/pre-step waterfall and appends a user-role recall message
|
|
4
|
+
* carrying the retrieved memories.
|
|
5
|
+
*
|
|
6
|
+
* Message construction MUST go through createUserMessage: the loop reads
|
|
7
|
+
* `decision.kind` and downstream consumers read `message.source.kind`, so a
|
|
8
|
+
* hand-rolled `{ role, content }` object aborts the turn with
|
|
9
|
+
* "Cannot read properties of undefined (reading 'kind')" (incident 2026-10-10).
|
|
10
|
+
* The reference implementation is the official dsh-time-context plugin.
|
|
11
|
+
*
|
|
12
|
+
* Memory is an OPTIONAL dependency (ctx.get) by design — this plugin must not
|
|
13
|
+
* sit PENDING when no provider is loaded; it just stays passive.
|
|
14
|
+
*/
|
|
15
|
+
|
|
16
|
+
import { createUserMessage } from '@deepseek-ai/dsh-llm'
|
|
17
|
+
|
|
18
|
+
import { MemoryInjector, defaultInjectConfig, } from './inject.js'
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
|
|
22
|
+
|
|
23
|
+
|
|
24
|
+
|
|
25
|
+
export const name = 'dsh-memory-inject'
|
|
26
|
+
|
|
27
|
+
export function apply(ctx , config = {}) {
|
|
28
|
+
const cfg = {
|
|
29
|
+
enabled: config.enabled ?? defaultInjectConfig.enabled,
|
|
30
|
+
tokenBudget: config.tokenBudget ?? defaultInjectConfig.tokenBudget,
|
|
31
|
+
topK: config.topK ?? defaultInjectConfig.topK,
|
|
32
|
+
}
|
|
33
|
+
|
|
34
|
+
// One injector per activation: its repeat-suppression state must live across
|
|
35
|
+
// the steps of a turn (a per-step instance would re-inject every step).
|
|
36
|
+
const injector = new MemoryInjector(
|
|
37
|
+
async (query, signal) => {
|
|
38
|
+
const memory = ctx.get('memory')
|
|
39
|
+
if (!memory?.search) return []
|
|
40
|
+
return memory.search({ text: query, signal, topK: cfg.topK })
|
|
41
|
+
},
|
|
42
|
+
cfg,
|
|
43
|
+
)
|
|
44
|
+
|
|
45
|
+
ctx.on('agent/pre-step', async (input, next) => {
|
|
46
|
+
// Runs outermost (prepend), so `next()` yields the fully assembled decision.
|
|
47
|
+
const decision = await next()
|
|
48
|
+
// A rejected step must pass through untouched — never rebuild the decision.
|
|
49
|
+
if (!decision || decision.kind === 'reject' || input.signal?.aborted) return decision
|
|
50
|
+
// optional capability: absence (or partial load) → pass through untouched
|
|
51
|
+
if (!ctx.get('memory')) return decision
|
|
52
|
+
|
|
53
|
+
const text = await injector.maybeInject(
|
|
54
|
+
{ messages: input.messages , signal: input.signal },
|
|
55
|
+
{ messages: decision.messages },
|
|
56
|
+
)
|
|
57
|
+
if (!text) return decision
|
|
58
|
+
const message = createUserMessage({
|
|
59
|
+
content: [{ type: 'text', text }],
|
|
60
|
+
source: { kind: 'memory', form: 'recall' },
|
|
61
|
+
})
|
|
62
|
+
// Last-resort guard: a malformed message aborts the whole turn downstream
|
|
63
|
+
// (consumers read message.source.kind), so never publish one.
|
|
64
|
+
if (!isWellFormedMessage(message)) return decision
|
|
65
|
+
return { ...decision, messages: [...decision.messages, message] }
|
|
66
|
+
}, { prepend: true })
|
|
67
|
+
}
|
|
68
|
+
|
|
69
|
+
const VALID_BLOCK_TYPES = new Set(['text', 'reasoning', 'image', 'file', 'tool-call', 'tool-addition', 'tool-removal'])
|
|
70
|
+
|
|
71
|
+
/** Structural check against the host Message contract (id/role/content/source.kind). */
|
|
72
|
+
function isWellFormedMessage(message ) {
|
|
73
|
+
const m = message
|
|
74
|
+
if (!m || typeof m.id !== 'string' || typeof m.role !== 'string') return false
|
|
75
|
+
if (!Array.isArray(m.content) || m.content.length === 0) return false
|
|
76
|
+
if (!m.content.every((block) => typeof (block )?.type === 'string' && VALID_BLOCK_TYPES.has((block ).type))) return false
|
|
77
|
+
const source = m.source
|
|
78
|
+
return typeof source?.kind === 'string'
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
//# sourceURL=D:\.deepseek\memory-plugin\packages\memory-inject\src\apply.ts
|
|
@@ -0,0 +1,127 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Memory injection logic (M3) — pure functions, no cordis imports.
|
|
3
|
+
*
|
|
4
|
+
* Injection shape follows the official docs: a user-role message whose content
|
|
5
|
+
* is a single `context` block with `contextForm: 'recall'`, prepended to the
|
|
6
|
+
* messages that enter the step via the agent/pre-step waterfall.
|
|
7
|
+
*/
|
|
8
|
+
|
|
9
|
+
import { estimateTokens } from '../provider/rerank.js'
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
|
|
19
|
+
export const defaultInjectConfig = {
|
|
20
|
+
enabled: true,
|
|
21
|
+
tokenBudget: 500,
|
|
22
|
+
topK: 5,
|
|
23
|
+
}
|
|
24
|
+
|
|
25
|
+
/** Minimal structural view of the messages we inspect (role + text blocks). */
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
|
|
34
|
+
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
|
|
38
|
+
|
|
39
|
+
|
|
40
|
+
/**
|
|
41
|
+
* Extract the retrieval query: the text of the LAST user message. Synthetic
|
|
42
|
+
* context injections (file notices etc.) are user-role too, so scanning from
|
|
43
|
+
* the end finds the freshest human-visible input; empty text → no injection.
|
|
44
|
+
*/
|
|
45
|
+
export function pickQueryText(messages ) {
|
|
46
|
+
for (let i = messages.length - 1; i >= 0; i--) {
|
|
47
|
+
const m = messages[i]
|
|
48
|
+
if (m.role !== 'user') continue
|
|
49
|
+
const text = (m.content ?? [])
|
|
50
|
+
.filter((b) => b.type === 'text' && typeof b.text === 'string')
|
|
51
|
+
.map((b) => b.text )
|
|
52
|
+
.join(' ')
|
|
53
|
+
.trim()
|
|
54
|
+
if (text) return text
|
|
55
|
+
}
|
|
56
|
+
return ''
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
/** Render hits into a recall text block, greedily within the token budget. */
|
|
60
|
+
export function renderRecallBlock(hits , cfg ) {
|
|
61
|
+
if (hits.length === 0) return null
|
|
62
|
+
const header = '以下是与此请求可能相关的长期记忆,供参考(可用 memory_search 核实,用 memory_forget 删除):'
|
|
63
|
+
let tokens = estimateTokens(header) + 4
|
|
64
|
+
const lines = []
|
|
65
|
+
for (const h of hits.slice(0, cfg.topK)) {
|
|
66
|
+
const line = `- ${h.record.content.replace(/\s+/g, ' ').slice(0, 300)}`
|
|
67
|
+
const cost = estimateTokens(line) + 1
|
|
68
|
+
if (tokens + cost > cfg.tokenBudget) break
|
|
69
|
+
lines.push(line)
|
|
70
|
+
tokens += cost
|
|
71
|
+
}
|
|
72
|
+
if (lines.length === 0) return null
|
|
73
|
+
return { text: `${header}\n${lines.join('\n')}`, tokens }
|
|
74
|
+
}
|
|
75
|
+
|
|
76
|
+
/** The context block we prepend (matches the documented ContextForm 'recall'). */
|
|
77
|
+
export function recallText(text ) {
|
|
78
|
+
return text
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
/**
|
|
82
|
+
* Turn-agnostic injector: searches memory for the step's query and returns the
|
|
83
|
+
* recall text to append. The host middleware (apply.ts) wraps that text into a
|
|
84
|
+
* real message via createUserMessage — a bare {role, content} object is NOT a
|
|
85
|
+
* message and crashes the loop (see the 2026-10-10 incident below).
|
|
86
|
+
*
|
|
87
|
+
* Repeated identical injections are suppressed so a multi-step turn does not
|
|
88
|
+
* repeat the same block before every model call.
|
|
89
|
+
*/
|
|
90
|
+
export class MemoryInjector {
|
|
91
|
+
lastText = ''
|
|
92
|
+
search
|
|
93
|
+
cfg
|
|
94
|
+
|
|
95
|
+
constructor(
|
|
96
|
+
search ,
|
|
97
|
+
cfg ,
|
|
98
|
+
) {
|
|
99
|
+
this.search = search
|
|
100
|
+
this.cfg = cfg
|
|
101
|
+
}
|
|
102
|
+
|
|
103
|
+
async maybeInject(
|
|
104
|
+
input ,
|
|
105
|
+
_decision ,
|
|
106
|
+
) {
|
|
107
|
+
if (!this.cfg.enabled) return undefined
|
|
108
|
+
const query = pickQueryText(input.messages)
|
|
109
|
+
if (!query) return undefined
|
|
110
|
+
|
|
111
|
+
let hits
|
|
112
|
+
try {
|
|
113
|
+
hits = await this.search(query, input.signal)
|
|
114
|
+
} catch {
|
|
115
|
+
return undefined // retrieval failure never breaks the agent loop
|
|
116
|
+
}
|
|
117
|
+
|
|
118
|
+
const block = renderRecallBlock(hits, this.cfg)
|
|
119
|
+
if (!block || block.text === this.lastText) return undefined
|
|
120
|
+
this.lastText = block.text
|
|
121
|
+
|
|
122
|
+
return block.text
|
|
123
|
+
}
|
|
124
|
+
}
|
|
125
|
+
|
|
126
|
+
|
|
127
|
+
//# sourceURL=D:\.deepseek\memory-plugin\packages\memory-inject\src\inject.ts
|
|
@@ -0,0 +1,108 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cordis wiring for dsh-memory-local. Loaded only inside DSH.
|
|
3
|
+
*
|
|
4
|
+
* Per the official docs:
|
|
5
|
+
* - registers the service so consumers reach it as `ctx.memory`
|
|
6
|
+
* - the LLM rerank caller adapts `ctx.llm` (one-shot GenerateOptions with
|
|
7
|
+
* RequestUserInput) — attached lazily via ctx.get, so the plugin never sits
|
|
8
|
+
* PENDING just because no LLM adapter is loaded
|
|
9
|
+
*/
|
|
10
|
+
import { Service, } from '@deepseek-ai/cordis'
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
import { LocalMemoryService, } from './index.js'
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
export const name = 'dsh-memory-local'
|
|
22
|
+
|
|
23
|
+
export class CordisMemoryService extends Service {
|
|
24
|
+
impl
|
|
25
|
+
|
|
26
|
+
constructor(ctx , config = {}) {
|
|
27
|
+
super(ctx, 'memory')
|
|
28
|
+
this.impl = new LocalMemoryService(config)
|
|
29
|
+
ctx.effect(() => () => this.impl.flushSync())
|
|
30
|
+
this.attachRerank(ctx, config)
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
/**
|
|
34
|
+
* Wire the rerank stage to ctx.llm when both it and a model are configured.
|
|
35
|
+
* Uses the documented one-shot shape: GenerateOptions + RequestUserInput,
|
|
36
|
+
* consuming `text-delta` chunks until finish.
|
|
37
|
+
*/
|
|
38
|
+
attachRerank(ctx , config ) {
|
|
39
|
+
const llmCfg = config.retrieval?.llm
|
|
40
|
+
if (!llmCfg?.enabled || !llmCfg.model) return
|
|
41
|
+
const llm = ctx.get('llm')
|
|
42
|
+
|
|
43
|
+
|
|
44
|
+
if (!llm?.stream) return
|
|
45
|
+
this.impl.setLlmInvoke(async ({ system, prompt, signal }) => {
|
|
46
|
+
const chunks = []
|
|
47
|
+
const stream = llm.stream({
|
|
48
|
+
provider: llmCfg.provider || 'deepseek',
|
|
49
|
+
model: llmCfg.model,
|
|
50
|
+
system,
|
|
51
|
+
messages: [{ role: 'user', content: [{ type: 'text', text: prompt }] }],
|
|
52
|
+
maxTokens: 300,
|
|
53
|
+
temperature: 0,
|
|
54
|
+
signal,
|
|
55
|
+
})
|
|
56
|
+
for await (const chunk of stream) {
|
|
57
|
+
if (chunk.type === 'text-delta' && typeof chunk.textDelta === 'string') chunks.push(chunk.textDelta)
|
|
58
|
+
}
|
|
59
|
+
return chunks.join('')
|
|
60
|
+
})
|
|
61
|
+
}
|
|
62
|
+
|
|
63
|
+
save(input ) {
|
|
64
|
+
return this.impl.save(input)
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
/** Direct store access for server-side consumers (e.g. the memory panel). */
|
|
68
|
+
get store() {
|
|
69
|
+
return this.impl.store
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
/** Embedding-backend lifecycle, exposed for the memory panel's buttons. */
|
|
73
|
+
embedServerStatus() {
|
|
74
|
+
return this.impl.embedServerStatus()
|
|
75
|
+
}
|
|
76
|
+
|
|
77
|
+
startEmbedServerManually() {
|
|
78
|
+
return this.impl.startEmbedServerManually()
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
stopEmbedServerManually() {
|
|
82
|
+
return this.impl.stopEmbedServerManually()
|
|
83
|
+
}
|
|
84
|
+
|
|
85
|
+
search(query ) {
|
|
86
|
+
return this.impl.search(query)
|
|
87
|
+
}
|
|
88
|
+
|
|
89
|
+
forget(input ) {
|
|
90
|
+
return this.impl.forget(input)
|
|
91
|
+
}
|
|
92
|
+
|
|
93
|
+
list(input ) {
|
|
94
|
+
return this.impl.list(input)
|
|
95
|
+
}
|
|
96
|
+
}
|
|
97
|
+
|
|
98
|
+
export function apply(ctx , config = {}) {
|
|
99
|
+
class ScopedMemoryService extends CordisMemoryService {
|
|
100
|
+
constructor(ctx2 ) {
|
|
101
|
+
super(ctx2, config)
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
ctx.plugin(ScopedMemoryService)
|
|
105
|
+
}
|
|
106
|
+
|
|
107
|
+
|
|
108
|
+
//# sourceURL=D:\.deepseek\memory-plugin\packages\memory-local\src\apply.ts
|