Memory and knowledge · DeepSeek Harness plugins — Page 30

A persistent memory plugin for DeepSeek Harness that enables agents to store and recall information across sessions

DSH 归档增强插件：自动定期归档、文件夹归档整理、LLM 摘要沉淀经验库、会话收藏与便签，查看/恢复归档会话与一键关闭 dsh（恢复与关闭需应用核心补丁）

AutoMemory for DSH：会话级记忆（收尾提醒 + 软上限检查 + 记忆索引维护），配置面板在 设置→插件。

组装式上下文（Assembled Context）: a context tree over the dsh session surface with per-node assemble modes (full / key / off), agent-authored presets, and a browser panel that decides what the model actually sees.

DSH Agent 跨会话全局记忆插件：会话开始注入条目级索引，模型按需 recall 记忆全文，显式保存/搜索/删除，数据仅存本机 $DSH_HOME/memory/

Jev decision toolkit for DeepSeek Harness: 21 named, typed judgments (privacy scan, change-scope, memory triage, batch log/bug/alert triage, agent-loop checks) built on one hardened, budgeted, ledger-backed transport. Advisory only — it never blocks.

Local-first RAG knowledge tools for DSH: kb_query / kb_ingest / kb_crawl / kb_related, backed by the kb-rag Python pipeline (Ollama bge-m3 + ChromaDB + SiYuan). Zero API cost, data never leaves your machine.

Persistent cross-session local memory for DSH with Markdown SSOT and self-healing SQLite FTS mirror

Codex-like persistent memory for DeepSeek Harness: auto-injected memory summary, raw memory tools, and per-session rollout summaries distilled into a bounded global memory file.

Qoder-Agent-SDK-style persistent memory for DSH: a background generation pass writes knowledge into per-scope memory files after each turn, and a consumption pass loads them into Agent context on session start or explicit refresh. DSH 持久记忆插件：按 Qoder SDK 的做法分离「生成」与「消费」，user/project 双作用域，token 预算加载。

DSH 记忆管理插件（宿主级）：文件夹记忆库、印象、组合与双向链接、上下文计划注入、规约记忆（新会话自动常驻）、Agent 记忆工具、图谱视图，设置页自由开关。

忠实复刻 oh-my-openagent 核心能力：ultrawork 协议 / rules 引擎 / hashline 精编 / boulder 记忆 / 结构化代码搜索 / 评论检查 / 后台监控 / 交接 / 团队任务表 / OmO 控制台面板

DSH 角色设定与长期记忆：角色库（persona.yml + SOUL.md + USER.md）、全局/项目两级记忆（FACT.md / MEMORY.md）、长期记忆 JOURNAL.jsonl；首条消息快照注入，对所有 agent 预设生效，附右侧栏管理面板

项目记忆（Project Memory）：本地优先的项目记忆插件。每个项目一份 .dsh-memory.json（按 cwd 隔离），提供 memory_read/recall/write/update/delete 五工具、/memory 命令、设置页记忆管理 UI（编辑/删除/拖动排序/一键保存排序/状态过滤/搜索/多项目切换），全部遵循'随核心快速适配'（适配器+探测-退化-上报）与'万物皆插件'（可独立挂载/停用）。【单线 0.1.3】2026-09-11 起取消双线：只面向 DSH 最新核心（0.1.5-rc.1 及更新，桌面端 2.0.7+/2.0.9），不再为旧核做适配；含读取保序修复与「保存排序」二次确认/回读校验。同属 0.1.x 系列——**只有出现较大优化或新功能时才升 0.2.0**（本版暂不发布）。

Auditable, workspace-local project wiki for DeepSeek Harness with a native Web view

A workspace-scoped DSH knowledge repository with explicit immutable local file backup.

DSH plugin: architecture-first project governance. v0.10.0 makes the model a TREE + GRAPH — the containment tree (who belongs to whom) now sits beside a dependency graph (who imports whom), and the graph is DERIVED from a static import scan: never registered, never hand-written, never stale. Seven gates (anchor / scope / mainline / impact / repeat-patch / decision / completion) are pure queries inside the single write entry point nav_commit, so there is no begin/done lifecycle and no orphan state: closure is decided by EVIDENCE (did the scope files change?), never by session identity — kill a session and its intent still closes (A1). Governance surface is minimal by construction: anything that would rot when code changes must be a PROJECTION, not a document — hand-written surface is two contracts (ARCHITECTURE.md / AGENTS.md), and .internal/arch/*.md dropped its sha1-fingerprint/freshness machinery to become an on-demand projection. ONE append-only event log (.internal/events.jsonl) remains the single source of truth; the runtime model (incl. the dependency graph) is disposable — delete runtime/ and governance loses nothing (I3). Tool face 6: nav_graph (read, incl. mode=impact) / nav_commit (intent) / nav_decide (ADR) / nav_node (nodes) / nav_render (projections) / nav_set (mainline vector).

小双六层记忆引擎 TypeScript 实现（core/system/important/timeline 分层，格式与桌面 XiaoShuang 兼容）

小双人格插件 for DeepSeek Harness：六层记忆工具(记住/回忆/删除/归档) + 小双人设自动注入 + mood API + 素材路由 + TTS 语音播报

Persistent memory system plugin for DeepSeek Harness: atomic facts, recall/remember/consolidate engines, soul.md & user.md, and a browser settings section.

ego lite for DeepSeek Harness, with a memory it keeps: the agent recalls what it learned about a site before it acts, promotes a script that worked into a reusable site tool, and hands you the keyboard through a real dsh prompt when a page needs a human.

dsh plugin: SM3 tamper-evident execution audit chain (port of eco-agent trace_audit.py)

dsh plugin: scored memory tree with BM25 search and Obsidian bidirectional sync (port of eco-agent memory_tree.py)

dsh plugin: L1-L4 risk permission gate (port of eco-agent permissions.py)