dsh-memory-plus

仓库内包含的插件

• dsh-memory-bundlev0.1.0✓

  Meta-bundle: installs all dsh-memory plugins (CJK search, dedup, hybrid memory index, memory tool, compaction locator, core memory, skill manager + self-evolution) into a DeepSeek Harness profile as one patch layer.

  package.json

  packages/dsh-memory-bundle/package.json

  dsh.bundle.patch

  packages/dsh-memory-bundle/cordis.patch.yml

• dsh-compaction-locatorv0.1.0

  Near-lossless compaction for the DeepSeek Harness — extends BasicCompactionEngine so each <compacted-summary> ends with an Exact Sources locator block (spill files, touched file paths, seq range), making compacted details recoverable via memory_search / read.

  package.json

  packages/dsh-compaction-locator/package.json

  dsh.bundle.patch

  —

  官方核心包 (6)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-compaction^0.1.0-rc.7共享
  • @deepseek-ai/dsh-compaction-basic^0.1.0-rc.7共享
  • @deepseek-ai/dsh-llm^0.1.0-rc.7共享
  • @deepseek-ai/dsh-session^0.1.0-rc.7共享
  • @deepseek-ai/dsh-token-meter^0.1.0-rc.7共享
• dsh-memory-corev0.1.0

  Cross-session core memory for the DeepSeek Harness — workspace-scoped persistent facts (ctx.memoryCore) with a stable system-prompt section (KV-cache safe) and a memory_remember tool.

  package.json

  packages/dsh-memory-core/package.json

  dsh.bundle.patch

  —

  官方核心包 (4)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-system-prompt^0.1.0-rc.7共享
  • @deepseek-ai/dsh-tools^0.1.0-rc.7共享
  • @deepseek-ai/schemastery^3.18.1私有副本
• dsh-memory-indexv0.1.0

  Hybrid memory search for the DeepSeek Harness — sqlite-vec vector arm + FTS5 lexical arm (RRF fusion) over session events, with local bge embeddings via transformers.js. Provides ctx.memorySearch.

  package.json

  packages/dsh-memory-index/package.json

  dsh.bundle.patch

  —

  官方核心包 (4)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-session^0.1.0-rc.7共享
  • @deepseek-ai/dsh-session-query^0.1.0-rc.7共享
  • @deepseek-ai/schemastery^3.18.1私有副本
• dsh-memory-skillsv0.1.0

  Skill manager + background self-evolution for DeepSeek Harness — model-facing skill_write/delete/list tools persist DSH-native skill files, and a timer-driven reflection loop distills reusable skills from finished agent turns (background, KV-free, logged).

  package.json

  packages/dsh-memory-skills/package.json

  dsh.bundle.patch

  —

  官方核心包 (5)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-llm^0.1.0-rc.7共享
  • @deepseek-ai/dsh-skill^0.1.0-rc.7共享
  • @deepseek-ai/dsh-tools^0.1.0-rc.7共享
  • @deepseek-ai/schemastery^3.18.1共享
• dsh-memory-toolv0.1.0

  Model-facing memory_search tool for the DeepSeek Harness — hybrid lexical+semantic recall over a session's earlier conversation via ctx.memorySearch, with bounded snippets.

  package.json

  packages/dsh-memory-tool/package.json

  dsh.bundle.patch

  —

  官方核心包 (3)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-tools^0.1.0-rc.7共享
  • @deepseek-ai/schemastery^3.18.1私有副本
• dsh-session-query-sqlite-cjkv0.2.0

  CJK-aware ctx.sessionQuery backend for the DeepSeek Harness — SQLite FTS5 provider with a trigram tokenizer table so Chinese full-text search actually matches (upstream unicode61 does not segment CJK), plus a LIKE fallback for 1-2 character CJK queries.

  package.json

  packages/dsh-session-query-sqlite-cjk/package.json

  dsh.bundle.patch

  —

  官方核心包 (5)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-session^0.1.0-rc.7共享
  • @deepseek-ai/dsh-session-persistence^0.1.0-rc.7共享
  • @deepseek-ai/dsh-session-query^0.1.0-rc.7共享
  • @deepseek-ai/schemastery^3.18.1私有副本
• dsh-tool-result-dedupv0.1.0

  Tool-result dedup policy for the DeepSeek Harness — replaces byte-identical repeated plain-text tool results with a pointer to the first occurrence (hash-indexed), saving input tokens on long sessions.

  package.json

  packages/dsh-tool-result-dedup/package.json

  dsh.bundle.patch

  —

  官方核心包 (5)

  • @deepseek-ai/cordis^4.0.1共享
  • @deepseek-ai/dsh-output-retention^0.1.0-rc.7共享
  • @deepseek-ai/dsh-session^0.1.0-rc.7共享
  • @deepseek-ai/dsh-tools^0.1.0-rc.7共享
  • @deepseek-ai/schemastery^3.18.1私有副本

路径固定于提交 5a66aea598f9

概览

可以有效减少token消耗！DeepSeek Harness memory plugins (dsh-plugin): CJK-aware session full-text search, tool-result dedup, hybrid vector+FTS5 memory search, cross-session core memory, near-lossless compaction locators. Real-harness verified, 65 tests.