dsh-goodmemory
为 DeepSeek Harness 自动进行跨会话 GoodMemory 召回和回写
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:hjqcan/dsh-goodmemory#5301beb7e107a3822554fc1d06d01d8dceb13e5cGoodmemory 以 dsh-goodmemory 发布，当前版本为 0.1.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Automatic, durable, cross-session memory for DeepSeek Harness, backed by GoodMemory.
The plugin recalls relevant memory before a DSH model step and writes the completed human/assistant exchange after a successful turn. Recalled context is an identified DSH user/message, so everything visible to the model remains reconstructable from the session log.
completed DSH turn -> GoodMemory remember -> later DSH pre-step
-> GoodMemory recall -> durable user/message -> model
| Package/runtime | Supported version |
|---|---|
@deepseek-ai/dsh | 0.1.0-rc.8 |
goodmemory | 0.8.3 |
| Node.js | ^22.19.0 or >=24.0.0 |
| Bun, managed mode only | >=1.3.14 |
DSH is still an RC. This package deliberately pins its DSH peers; a new RC is unsupported until the integration suite passes against it.
Install from npm:
dsh plugin --profile web add dsh-goodmemory@0.1.2
dsh --profile web --dump-config
dsh --profile web
For a local checkout:
pnpm install
pnpm pack
dsh plugin --profile web add ./dsh-goodmemory-0.1.2.tgz
Use --profile headless instead of web for the headless composition. The bundle inserts one goodmemory row; later profile, home, and CLI patch layers can replace its complete config.
Managed mode is the default. It starts the published goodmemory-http-bridge on a random loopback port, creates an in-memory bearer token, enables the recommended retrieval preset, and stores SQLite data at $DSH_HOME/goodmemory/memory.sqlite.
Install Bun first. GOODMEMORY_BUN_BINARY may name a non-default Bun executable. To choose a different database path, override the row in the profile's cordis.patch.yml:
- id: goodmemory
name: dsh-goodmemory
config:
mode: managed
databasePath: /absolute/path/to/memory.sqlite
The plugin waits for in-flight writeback during normal HMR or process teardown, then terminates the managed sidecar. It does not automatically restart a sidecar that crashes after activation; DSH keeps running without memory and logs the failure until the plugin is reloaded.
External mode connects to an existing GoodMemory HTTP bridge. Keep the token in an environment variable, not in DSH config:
export GOODMEMORY_HTTP_BRIDGE_TOKEN='replace-with-a-secret'
goodmemory-http-bridge --recommended
- id: goodmemory
name: dsh-goodmemory
config:
mode: external
baseUrl: http://127.0.0.1:8739
tokenEnv: GOODMEMORY_HTTP_BRIDGE_TOKEN
Activation verifies health, the exact phase-39.http-memory.v1 contract, and authenticated recall. Invalid config, a missing token, failed authorization, or a contract mismatch fails plugin activation.
type Config = {
recall?: boolean // default true
writeback?: boolean // default true
maxRecallTokens?: number // default 256
requestTimeoutMs?: number // default 10000
scope?: {
userId?: string
workspaceId?: string | null
agentId?: string | null
}
} & (
| { mode?: 'managed'; databasePath?: string; startupTimeoutMs?: number }
| { mode: 'external'; baseUrl: string; tokenEnv?: string }
)
agent/pre-step batch contains direct source.kind === "user" text. Tool-only continuations and plugin context do not trigger another lookup.With GoodMemory 0.8.3, a direct user statement such as the following is accepted without annotations and recalled after a process restart in the same scope:
Project decision: When SQLite reports SQLITE_BUSY_SNAPSHOT, roll back the transaction, begin again and recompute from a fresh read before writing.
We decided that ... and substantive Project policy: ... declarations also
work. Questions and undecided placeholders are not confirmed decisions.
GoodMemory 0.8.2 retained the example above as source text but did not create a
durable fact. Re-send previously source-only decisions after upgrading; the
upgrade does not re-extract old transcripts automatically.
Every completed writeback logs a writeback_result receipt with the session,
turn, duration, accepted/rejected counts, outcome, extraction strategy, and
known rejection reason codes. no_admissible_candidate means no durable
candidate was admitted; HTTP success alone does not mean the text became a
memory. These diagnostics omit conversation text, raw bridge errors, and
credentials. Recall counts are available through debug-level recall_result
logs. Extraction warnings or a failed outcome produce warning-level receipts.
Defaults are deliberately DSH-isolated:
userId: DSH's persisted anonymous user id under $DSH_HOME;workspaceId: SHA-256 of the normalized session working directory, or dsh-workspace:global without a working directory;agentId: dsh;sessionId: the current DSH session id.GoodMemory's durable scope excludes sessionId, so different DSH sessions in the same user/workspace/agent scope share long-term memory. To share with another GoodMemory host, explicitly configure the same userId, workspaceId, and agentId; set nullable workspace/agent fields to null when the other host omits them.
The plugin does not copy DSH JSONL logs or create a second memory schema. It sends only selected text to GoodMemory: direct human text plus the final assistant text. GoodMemory remains the sole memory/storage truth.
The first release intentionally adds no model tools, MCP server, DSH command, or management UI. Use GoodMemory's existing CLI and Inspector for health, review, revision, export, and forgetting:
goodmemory inspector serve
pnpm test
pnpm typecheck
pnpm build
pnpm test:e2e # real Bun + GoodMemory SQLite restart proof
pnpm pack
The deterministic integration suite uses a scripted DSH model adapter. It verifies logged recall context, completed-only writeback, tool/reasoning exclusion, scope isolation, write-after-read ordering, runtime degradation, and strict startup validation. The decision regression runs write, restart, and different-directory phases in separate Node processes with a real managed Bun/SQLite bridge, inspecting both the model request and session log. CI also runs that regression on Windows. No benchmark uplift or answer-quality claim is made.