chenzheshushi-commits/dsh-evolve ↗★ 0
dsh-evolve
Self-evolving memory + skill lifecycle for DeepSeek Harness. Cross-session memory with zero-token deterministic recall (bigram-Jaccard fused with FTS5 BM25 via RRF), a tiered approval gate, and reinforcement that strengthens what you repeat. Procedural knowledge crystallizes into SKILL.md files that refine in place and are curated through an active-stale-archived lifecycle (reversible archive, pre-op backups, rollback, never deletes). Includes background per-turn review, anti-bloat convergence for both skills and memory, an auto-grown user profile, and a web settings page.
安装
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:chenzheshushi-commits/dsh-evolve说明文档
阅读完整 README ↗Configuration
Everything is configurable through the plugin's settings page (web profile) or your DSH config. Notable switches:
| Key | Default | Effect |
|---|---|---|
autoConfirmEnabled | true | false = every model write waits for review |
reviewEnabled | true | Background per-turn review |
reviewEveryTurns | 5 | Review throttle |
reviewModel | (main model) | Route review to a different model |
refineLLM | false | Use an LLM pass when crystallizing/refining skills |
reinforceEvery | 3 | Observations per importance step |
memoryMaxChars | 20000 | Memory character budget (0 disables) |
convergeSuggest | true | Surface merge/fold suggestions |
curatorStaleDays / curatorArchiveDays | 30 / 60 | Skill lifecycle thresholds |
ftsEnabled | true | false = pure bigram recall, no SQLite |
The LLM is only ever used for optional auxiliary passes — skill refinement, background review, and skill merging. All of them are single-shot, skippable, and fall back to deterministic behavior on failure. Nothing runs in your main loop.