dsh-memory-setup
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。
Install
dsh plugin --profile <profile> add dsh-memory-setup
Tools
| 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 |
Storage & auditability
- Location:
<workspace>/.dsh-memory-setup/memory.json — plain JSON, easy to read/back up
- Every mutation appends to
changelog (when / what / why) — memory is auditable by design
- Lessons carry an optional
evidence field (file/command/observation) — no evidence, no lesson
- Local-first: nothing leaves your machine
Config (optional)
| 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 |
Roadmap
- v0.2 ✅: incident review with dedupe (
memory_review), lesson/convention expiry + changelog cap (auto-pruned on save), memory.json.bak backup on every save, Markdown export (memory_export)
- v0.3 ✅: personal knowledge base (
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 fallback
- v0.4 ✅: BM25 retrieval for the knowledge base (title ×3 / tags ×2 / content ×1, IDF-scaled — no embeddings, no deps), memory diff export (
memory_diff vs backup), self-review reminder in the injected guidance
- v0.5 ✅: optional embeddings provider (OpenAI-compatible endpoint;
knowledge_embed backfill + cosine retrieval, BM25 fallback), bulk incident review (memory_review_session), own-tool fs failure tracking surfaced into the injected context
- v0.6 ✅: memory snapshots & restore (
memory_snapshot / memory_list_snapshots / memory_restore, capped, index-file based), fault troubleshooting (memory_troubleshoot — lessons + knowledge lookup), troubleshoot reminder in guidance
- v0.7 ✅: KB included in snapshots (snapshot/restore both memory + knowledge), lesson auto-promotion (recurring lessons with hits ≥ threshold become standing conventions, auto-run on save — the memory literally learns from repeated mistakes), memory stats (
memory_stats)
- v0.8 ✅: schema v2 migration (auto on load), memory import/merge (
memory_import/memory_merge, newer/both conflicts), knowledge base JSON round-trip (kb_export/kb_import)
- v0.9 ✅: lesson health evaluation (failing/resolved/active — auto on save, ⚠️ markers in render), relevance-focused injection (
memory_focus)
- v1.0 ✅: tiers (hot/warm/cold), integrity audit (
memory_audit, sha256), privacy redaction in exports (sensitiveKeys)
- v1.1 ✅: optimistic locking (revision-based CAS — multi-session/multi-agent safe), CLAUDE.md import (
memory_import_claude)
- v1.2 ✅: full backup bundle (/ — memory + KB + snapshots), (, owner/purpose) — team-ready
Security
Memory plugins are the highest-trust plugin type — see SECURITY.md for the audit posture.
