Skip to content

Releases

v0.1.1 — memory and self-evolving skills for DSH

Choose a tag to compare

@Ricky74192 Ricky74192 released this 29 Oct 11:19

dsh-memory-forge gives DeepSeek Harness the memory loop it does not ship with: automatic turn capture, recall injected before each step, four memory tools, and skill distillation into real SKILL.md bundles.

Install

dsh plugin --profile <profile> add git+https://github.com/Ricky74192/dsh-memory-forge.git
# or, without git installed:
dsh plugin --profile <profile> add https://github.com/Ricky74192/dsh-memory-forge/archive/refs/heads/main.tar.gz

Then restart the profile, or let HMR reload it. The four tools appear in every session and capture/recall start immediately — no YAML required.

Inside

  • memory_save, memory_search, memory_forget, memory_distill_skill
  • append-only JSONL under $DSH_HOME/memory, skills under $DSH_HOME/skills
  • BM25 + query coverage + recency, CJK bigrams, function-word filtering
  • zero dependencies, no network calls, no native modules

Verification

  • 62 unit tests
  • four-phase offline e2e (capture, recall injection, skill distillation, harness discovery) on three install paths
  • CI green on Node 22 and 24

Details: README.zh.md / README.md, data handling in SECURITY.md.

v0.1.0 — memory and self-evolving skills for DSH

Choose a tag to compare

@Ricky74192 Ricky74192 released this 29 Oct 09:27

dsh-memory gives DeepSeek Harness the memory loop it does not ship with: automatic turn capture, recall injected before each step, four memory tools, and skill distillation into real SKILL.md bundles.

Install

dsh plugin --profile <profile> add git+https://github.com/Ricky74192/dsh-memory.git
# or, without git installed:
dsh plugin --profile <profile> add https://github.com/Ricky74192/dsh-memory/archive/refs/heads/main.tar.gz

Then restart the profile, or let HMR reload it. The four tools appear in every session and capture/recall start immediately — no YAML required.

Inside

  • memory_save, memory_search, memory_forget, memory_distill_skill
  • append-only JSONL under $DSH_HOME/memory, skills under $DSH_HOME/skills
  • BM25 + query coverage + recency, CJK bigrams, function-word filtering
  • zero dependencies, no network calls, no native modules

Verification

  • 62 unit tests
  • four-phase offline e2e (capture, recall injection, skill distillation, harness discovery) on three install paths
  • CI green on Node 22 and 24

Details: README.zh.md / README.md, data handling in SECURITY.md.