dsh-file-memory
提供分层记忆管理与背景整合功能 适合需要工作区、角色文件及语义召回等原生记忆增强的用户。
安裝
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:aqsk-BLG/dsh-memory說明文件
閱讀完整 README ↗Configuration
| Key | Default | Meaning |
|---|---|---|
dshHome | unset | Advanced explicit override only. Normally omit it so the plugin uses the same root already selected by DSH: $DSH_HOME, otherwise DSH's default ~/.dsh |
versionGate | error | Host-version gate at load: error fails loud on hosts below 0.1.0-rc.7, warn logs instead, off disables the check |
memoryBudgetChars | 4000 | Code-point budget for global MEMORY.md |
userBudgetChars | 1500 | Code-point budget for global USER.md |
projectMemoryBudgetChars | 3000 | Code-point budget for live project MEMORY.md |
reminderEnabled | true | Include the skip-aware every-turn memory reminder |
memorySectionOrder | 5 | Frozen global-memory prompt order |
identityBudgetChars | 4000 | Code-point budget for IDENTITY.md |
soulBudgetChars | 4000 | Code-point budget for SOUL.md |
seedMissingPersonaFiles | true | Seed conservative persona defaults when absent |
seedMissingMemoryFiles | true | Seed short USER.md/MEMORY.md templates when absent; never overwrites |
personaSectionOrder | -50 | Persona-file prompt order before deployment persona |
maxHits | 20 | Maximum sessions one session_search call may return |
semanticEnabled | true | Use semantic ranking when a model route is available |
semanticProvider | "" | Optional dedicated semantic provider, paired with semanticModel |
semanticModel | "" | Optional dedicated semantic model, paired with semanticProvider |
semanticBatchSize | 30 | Candidates per semantic-ranking request |
semanticCandidateChars | 2000 | Code-point budget retained from each past session |
semanticMaxTokens | 2048 | Output-token cap for each ranking request |
semanticReadConcurrency | 4 | Concurrent past-session surface reads |
semanticFallbackEnabled | true | Use and label full-text fallback when semantics cannot run |
flushEnabled | true | Queue the reminder after successful compaction |
consolidationEnabled | true | Run skip-aware reviews after eligible completed turns |
consolidationMode | automatic | Write controlled regions or log proposal candidates only |
consolidationEveryEligibleTurns | 1 | Eligible completed tasks per ordinary review batch; raise above one only to opt into batching |
consolidationMinUserChars | 12 | Minimum human code points for a non-tool turn |
consolidationMinAssistantChars | 24 | Minimum assistant code points for a non-tool turn |
consolidationMaxTurnsPerReview | 20 | Maximum eligible turns in one review |
consolidationTranscriptBudgetChars | 12000 | Bounded review transcript text |
consolidationDailyBudgetChars | 1200 | Daily append budget per review |
consolidationMaxTokens | unset | Advanced route override only; normally omit it so the selected model adapter supplies its native output limit |
consolidationReasoningEffort | unset | Normally inherit the live session's effort on the same route, or the dedicated route's adapter default |
consolidationTimeoutMs | 180000 | End-to-end background-review deadline in milliseconds |
consolidationMaxDeletionRatio | 0.5 | Larger automatic managed-list deletions become proposals; a full clear is always protected unless explicitly requested |
consolidationRetryBaseDelayMs | 60000 | Initial transient-failure retry delay |
consolidationRetryMaxDelayMs | 3600000 | Maximum transient-failure retry delay |
consolidationProvider | "" | Optional dedicated review provider, paired with consolidationModel |
consolidationModel | "" | Optional dedicated review model, paired with consolidationProvider |
For a raw Cordis composition rather than a profile bundle:
- name: dsh-file-memory
config:
reminderEnabled: true
semanticEnabled: true
flushEnabled: true
consolidationEnabled: true