dsh-agent-memory：为 DSH 提供智能体驱动长期记忆
dsh-agent-memory 是为 DeepSeek Harness (DSH) 打造的 MIT 许可长期记忆插件（v0.3），旨在提供跨会话的结构化记忆能力。它采用分层存储架构，将记忆划分为 L1 事实（偏好与环境）、L2 知识（可复用内容）、L3 情景（经历时间线）及摘要层，并支持全局与工作区两种作用域路由。插件向智能体暴露 remember、recall、memory_browse、update/forget 及 memory_stats 等工具，支持多维度检索与时间金字塔浏览。其特色功能包括时间桶自动压缩以及与 dsh-agent-compact 联动，在会话压缩后自动沉淀 che
Read MoreDeepSeek Harness Plugin `dsh-negative-ledger`: Recording Falsified Paths and Automatically Invalidating Them Upon Evidence Changes
`@akslcw/dsh-negative-ledger` is a community plugin for DeepSeek Harness (DSH), designed to persistently record an agent's "falsified paths" and prevent repeated invocation of failed commands, attempts to read missing files, or calls to unavailable APIs. This plugin is not a memory system, cache, or bug tracker; instead, it transforms failed attempts into traceable, invalidatable negative facts by binding evidence and retry conditions. Its core mechanisms include: identifying repeated attempts via fingerprints based on command or file path; offering three modes—`warn` (default, with added contextual hints), `block` (intercepts before execution), and `off`; and automatically invalidating conclusions when dependent evidence changes to allow retries. The plugin supports SQLite (multi-process/WAL) and JSONL (single-process) backends.
Read MoreImplementing Traceable Knowledge Graph Memory for DeepSeek Harness with graph-memory
Graph-memory is a DSH memory-type plugin maintained by adoresever, which extracts dialogues into TASK/SKILL/EVENT knowledge graphs, supporting cross-session semantic recall and local SQLite storage. This article introduces its core capabilities, directory page installation commands, source tarball installation steps, gm_* tools and vector retrieval configuration, as well as the plugin permissions, beta version, and the application boundaries of the 75% context compression benchmark.
Read More