dsh-memory-setup
解决 AI 金鱼脑问题：为 DeepSeek Harness 提供个人记忆层，用于存储偏好、项目约定、工作流程和错误经验，并以可审计、有证据支持的 Markdown/JSON 格式保存在本地。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:863683348/dsh-memory-setup#989e8a5e7ac32b0864f5a7c2167ca5e242d55f43Memory Setup 以 dsh-memory-setup 发布，当前版本为 1.2.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Solve the AI goldfish brain 🐠 — a local, auditable personal memory layer for DeepSeek Harness. Remembers your preferences, project conventions, workflows, and error lessons, and injects them back into every session.
解决 AI 的"金鱼脑"：本地、可审计的个人记忆层——偏好、项目约定、工作方式、纠错教训，会话间自动继承。
dsh plugin --profile <profile> add dsh-memory-setup
| Tool | What it does |
|---|---|
memory_setup | One-time onboarding: language, code style, tools, conventions, workflows |
memory_status | Read current memory + changelog (also auto-injected guidance at boot) |
memory_update | Update one memory path (e.g. preferences.codeStyle) with a changelog entry |
memory_project | Auto-extract project conventions from workspace files (README / package.json / configs), preview or apply |
memory_lesson | Record an error lesson (error → fix → evidence) so the same mistake is not repeated |
memory_review | v0.2 — formalize an incident into a lesson with root cause; similar lessons are auto-merged (dedupe + hit counter) |
memory_export | v0.2 — export the full memory + changelog to a Markdown file for review/backup |
knowledge_add | v0.3 — add a knowledge entry (title/content/tags/source); similar titles auto-merge |
knowledge_search | v0.3 — keyword retrieval (title ×3 / tags ×2 / content ×1 scoring) |
knowledge_list / knowledge_remove | v0.3 — browse / delete knowledge entries |
memory_diff | v0.4 — diff current memory against memory.json.bak, optionally written to memory-diff.md |
memory_review_session | v0.5 — bulk incident review: submit many failures at once, dedupe per item |
memory_snapshot / memory_list_snapshots / memory_restore | v0.6 — snapshot the memory (keeps N), list, and restore with auto-backup of the current state |
memory_troubleshoot | v0.6 — given an error, search past lessons + knowledge base for a known fix |
memory_stats | v0.7 — aggregate stats across memory, knowledge base and snapshots |
memory_promote | v0.7 — promote recurring lessons (hits ≥ threshold) into standing conventions; auto-runs on save |
memory_import / memory_merge | v0.8 — import memory from JSON (auto-migrate) / merge two memories (newer or both) |
kb_export / kb_import | v0.8 — knowledge base JSON round-trip |
memory_focus | v0.9 — relevance-based injection: only memory matching a topic is injected |
memory_tier | v1.0 — hot/warm/cold tiers (hot is injected, cold is archived) |
memory_audit | v1.0 — sha256 integrity check + changelog audit report |
memory_import_claude | v1.1 — import conventions from CLAUDE.md |
memory_export_all / memory_import_all | v1.2 — full backup bundle (memory + KB + snapshots) |
memory_annotate | v1.2 — owner/purpose annotations on entries |
knowledge_embed | v0.5 — backfill embeddings for KB entries (needs embeddingEndpoint); enables semantic search |
<workspace>/.dsh-memory-setup/memory.json — plain JSON, easy to read/back upchangelog (when / what / why) — memory is auditable by designevidence field (file/command/observation) — no evidence, no lesson| Field | Default | Description |
|---|---|---|
| memoryDir | .dsh-memory-setup | memory dir relative to the session workspace |
| injectOnBoot | true | inject live memory into the system prompt (dynamic context, refreshed on save) |
| maxMemoryChars | 6000 | cap for rendered memory text |
| lessonTtlDays | 90 | lessons expire after this many days (0 disables) |
| changelogCap | 100 | max changelog entries kept |
| backupOnSave | true | write memory.json.bak before every save |
| reviewReminder | true | append self-review reminder to guidance |
| embeddingEndpoint | (empty) | OpenAI-compatible embeddings endpoint (enables semantic KB search) |
| embeddingKey | (empty) | Bearer key for the embeddings endpoint |
| embeddingModel | text-embedding-3-small | embeddings model name |
| snapshotKeep | 10 | max memory snapshots kept |
| troubleshootReminder | true | append troubleshoot/snapshot reminder to guidance |
memory_review), lesson/convention expiry + changelog cap (auto-pruned on save), memory.json.bak backup on every save, Markdown export (memory_export)knowledge_*, keyword retrieval, title-merge dedupe); dynamic memory injection via a live systemPrompt.context() section — refreshed at boot (from the workspace path) and after every memory save (throttled 30s), with static guidance as fallbackmemory_diff vs backup), self-review reminder in the injected guidanceknowledge_embed backfill + cosine retrieval, BM25 fallback), bulk incident review (memory_review_session), own-tool fs failure tracking surfaced into the injected contextmemory_snapshot / memory_list_snapshots / memory_restore, capped, index-file based), fault troubleshooting (memory_troubleshoot — lessons + knowledge lookup), troubleshoot reminder in guidancememory_stats)memory_import/memory_merge, newer/both conflicts), knowledge base JSON round-trip (kb_export/kb_import)memory_focus)memory_audit, sha256), privacy redaction in exports (sensitiveKeys)memory_import_claude)Memory plugins are the highest-trust plugin type — see SECURITY.md for the audit posture.
memory_export_allmemory_import_allmemory_annotate