Memory and knowledge · DeepSeek Harness plugins — Page 41

Minimalist long-term memory for DeepSeek Harness, distilling pi-loom + pi-esr ideas: zero-LLM auto-capture, a symbolic [LOOM] index with progressive disclosure, and an ESR-lite evidence-closure protocol (esr_task/esr_close/esr_link). Hot path is model-free; storage on ctx.storageDomain; web memory viewer and config card through DSH's native settings slots.

Local file recall for DeepSeek Harness: it watches the folders you choose and keeps a SQLite index of what appeared — name, type, size, appeared-at, location and origin — so "where did I put that file?" becomes a single question. Every folder carries its own incremental anchor, and a folder added later is backfilled in full on its first scan. Search is SQL hard filtering plus multi-keyword hit weighting, with a live fallback scan when the index has no hit. Read-only: it never modifies, moves or deletes your files. 文件快速寻回:定期扫描你指定的文件夹,把新出现的文件(名字/类型/大小/出现时间/位置/来源)记进本地 SQLite 索引,忘了东西放哪问一句就能找回。每个文件夹各有自己的增量锚点,新加入的文件夹第一次会回填全部历史;检索用 SQL 硬过滤加多关键词加权,索引没命中时当场兜底实时扫描一遍。只读,绝不修改、移动或删除你的文件。

Keeps a long agent turn answerable to the user's actual request: a standing maintainer-document reminder, two pre-execute gates (read a same-basename precedent before editing infrastructure; say what a side-effecting step is for), an objective anchor that quotes the user's latest instruction plus an explicit precedence rule, and a nudge stage that corrects before it blocks. Details: (1) A standing system-prompt section that tells the model to read the workspace maintainer documents (plan.md / conventions.md / stack.md / state.md / maintainer/README.md) before every operation, so small-model context compaction cannot erase long-term project memory. (2) A `tools/pre-execute` precedent gate that denies a write/edit inside a guarded infrastructure area until a working same-basename precedent has been read. (3) A `tools/pre-execute` intent gate that polices the turn's side-effecting calls when the model never said what the step is for. (4) An OBJECTIVE ANCHOR: the user's latest instruction, quoted verbatim in a second prompt section with an explicit precedence rule (user's latest instruction > your own last stated plan > a lead you found yourself), so a long turn cannot silently redefine its own task. (5) A NUDGE stage that corrects before it blocks — the first unexplained action only injects a reminder the model reads at its next step, a drift reminder fires once per long turn, and only a repeated offence is denied. Plus a right-sidebar tab that views and edits those documents, and ten runtime knobs exposed in the harness settings page (设置 → 插件 → 插件配置).

Mana 作意：注意力门控 + JEV 筛选 + 注入审计

Mana 熏习：合并 + SOAR chunking（P1 骨架）

Mana（末那识）核心契约插件：S1 五类事件 + 领域模型 + mana_trace/inject_log schema + 存储开库器

Mana 种子衰减：艾宾浩斯衰减 + 间隔重复 + 归档（P1 骨架）

Mana JEV 适配层：判定、缓存、熔断、三级降级（本机 Ollama logprob 承载）

Mana 学习链：Hebbian 共激活 → 关联强度（L-03；不含 FSRS）

Mana LLM 通道：注册 mana 路由到宿主 ctx.llm，使 LLM 生成有唯一出口（L-01）

Mana 长时记忆：编码 + ACT-R 激活方程 + 三路混合检索（P1 骨架）

Mana 末那识·审：置信度 / 知识缺口 + 可移植 shoucang 判据阈值层（B5.1）

Mana 前五识：输入采集、分块、实体抽取

Mana Prompt 模板库：v10 §25 的 20 个 Prompt 常量（真源）

Mana 再巩固：去稳定化窗口（四类记忆窗口参数）+ 内容更新可追溯（L-02）

Mana 调度：SOAR 目标栈 + 三阶段循环（P1）

Mana UI：Client 半区面板（激活值热力图 / 目标栈树 / 认知轨迹时间线）+ Host 侧读服务与 mana_trace 的 ui/render 落点

Mana 末那识·我执：偏好 / 习惯 / 项目（P2 骨架 · B5.1）

Mana 向量适配：嵌入、纯 JS 余弦检索、RRF 融合

Mana 第六识：工作记忆、容量管理、上下文组装

Markdown folder long-term memory for DeepSeek Harness: one file per fact, an index line per memory, mounted into every request as a system prompt section

Small DeepSeek Harness memory plugin: the model writes markdown files under ~/.dsh/memory, the host injects them next session. No auto-extract, no vector DB.

Mem0 persistent memory for the DeepSeek Harness web profile — automatic recall injection, tidal-coalesced memory writes, and mem0_search/add/update/delete tools against a self-hosted Mem0 server (X-API-Key). Zero-intrusion bundle plugin; no dsh source changes.

DSH plugin: per-workspace long-term memory store (.dsh-memory/memory.json) plus session archive management (archive / list / full-text search over archived sessions). Ships 7 model tools and a Settings page.