dsh-somni
Sleep-consolidated memory for DeepSeek Harness: resident identity + intentions, action-time associative recall, episodic / semantic / prospective memory, and dream-time consolidation via a Python sidecar. 适合需要长期记忆和情境关联的用户，通过Python端侧实现记忆整合。
설치
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:liuqingman/dsh-somniConfiguration
| Key | Default | Meaning |
|---|---|---|
dataDir | /somni | memory/, skills/, models/ live here |
python | python3 | Interpreter for the sidecar |
capture.enabled / maxMessageChars | true / 4000 | Write session-logs; per-message cap |
inject.identity / intentions / discipline | true | Resident prompt sections (orders 100 / 130 / 1200) |
assoc.enabled / preStep / postExecute | true | Associative hints on the latest message / after tool calls |
assoc.threshold / gap / maxItems | 0.52 / 0.04 / 2 | Confidence gate |
tools.enabled | true | Register the memory tools (+ dream_now) |
dream.enabled | true | Scheduled consolidation |
dream.idleMs / maxIntervalMs / minPendingLogs | 15 min / 6 h / 1 | Start after this much quiet, or at least this often while idle |
dream.llm | host | host = harness model via ctx.llm; openai = sidecar calls an OpenAI-compatible API (dream.openai.{baseUrl,apiKeyEnv,model}) |
dream.maxIterations | 60 | ReAct step cap per dream |
embed.provider | local | local (ONNX) · http (OpenAI-compatible /embeddings, embed.baseUrl) · off |
logLevel | info | Sidecar log level (forwarded to the harness logger) |
Dreams are preemptible: if any agent starts a turn mid-dream the sidecar is
told to stop, so the user never waits on consolidation. The dream_now tool
forces one from the conversation.