通用自我反思记忆系统：为 AI 代理设计的三级本地记忆
介绍 SkillHub 社区技能 @user_6bedba2d/memorypor（self-improving v2.1.0）：在 ~/self-improving/ 用热/温/冷三层文件管理纠错、自我反思与重复模式，无需 API Key。说明 skillhub 安装命令、setup.sh 初始化、与 AGENTS.md/CLAUDE.md 集成，以及三次确认晋升与 stats/heartbeat 维护方式。
Read MoreDeepSeek Harness Local Memory Plugin
The DeepSeek Harness (DSH) memory plugin adopts a fully local architecture and stores data in JSON files, ensuring privacy without requiring an API Key. It supports long-term memory across conversations and projects, provides four tools such as `memory_save` and `memory_search`, and allows memories to be managed and retrieval weights adjusted in the settings page. Data is stored in the `~/.dsh/` directory, Chinese text is segmented into two-character combinations, and tags are automatically assigned based on project paths. Although the plugin cannot intercept native deletion actions, it addresses the issue of the AI forgetting agreements due to DSH's limited context window, while maintaining the advantage of running entirely locally.
Read Moredsh-second-brain: DeepSeek Harness Second Brain Loader
This plugin is designed to inject a local “brain” snapshot (identity, state, etc.) into every new DeepSeek Harness (DSH) session, enabling cross-session memory. It runs as a read-only, zero-dependency, pure ESM module, injecting the snapshot into the system prompt via hooks while enforcing a 300-character minimum and a 1000-character maximum to prevent overflow. It is enabled by default after installation; new sessions automatically include the snapshot content, and users can verify the effect with specific commands.
Read Moredsh-flymemory: Local Long-Term Memory for DeepSeek Harness
dsh-flymemory is a DeepSeek Harness plugin designed to address the challenges of limited context and difficult state persistence in local agent development. The plugin runs the FlyMemory service locally and provides 15 memory tools and two automatic hooks, supporting memory CRUD operations, state tracking, and conflict management. It ensures that data is stored only in the local file system and does not involve network transmission. Built with Python, the plugin relies on `torch` and `sentence-transformers` for vector retrieval. After installation, users must configure environment variables and download the model; initial startup is time-consuming, but the plugin supports offline operation.
Read Moreviya-memory: A Local Long-Term Memory Plugin for DeepSeek Harness
viya-memory is a local storage plugin for DeepSeek Harness (DSH), designed to provide long-term memory for agents through Markdown files. It is a zero-dependency, single-file plugin that supports materializing memories into cards and leverages Obsidian's graph view, bidirectional links, and search features. After installation, you need to configure the path to the Obsidian vault. The plugin provides seven core tools, including saving, searching, updating, linking, deleting, and statistics. It features atomic writes, overwrite prevention, and sensitive-content scanning, making it suitable for developers who require data privacy and local control.
Read Moredsh-motion-memory: Local Memory Management Plugin for DeepSeek Harness
dsh-motion-memory is a memory management plugin designed for DeepSeek Harness (DSH), aimed at solving the context loss problem in agent workflows. It leverages conversation tracking, periodic summarization, and a four-layer memory structure to automatically consolidate conversation content into local memory documents, while supporting reference traceability and read-only protection. The plugin fully supports local storage and local models, with an emphasis on privacy and cost control. Installation requires a specific command and strict pairing with DSH version 0.2.0-rc.2.
Read MoreDeepSeek Harness Plugin: MemoKnow
MemoKnow is a local personal memory and knowledge management plugin for DeepSeek Harness (DSH). It stores data using SQLite and supports document snapshots, full-text search, and semantic search to help agents maintain contextual memory. Users can import and manage various document formats, and review, edit, or filter memories in the DSH interface. The plugin runs entirely locally, does not support multi-user use, and provides flexible learning control options.
Read Moredsh-memory: Localized Long-Term Memory for DeepSeek Harness
`dsh-memory` is a plugin for DeepSeek Harness that addresses the problem of agents forgetting user preferences and project constraints after a regular session ends. It integrates memory organization approaches from Claude Code and Codex, providing core capabilities such as local Markdown storage, global and workspace isolation, progressive disclosure, and secure writes. Users can enable it with an installation command, configure the memory organization model in the Web interface, and achieve long-term memory management across projects. The plugin is open source under the MIT license, ensuring data privacy and a local-first approach.
Read Moredsh-local-memory: DeepSeek Harness Local Cross-Session Memory Plugin
dsh-local-memory is a local memory persistence plugin designed for DeepSeek Harness. By automatically capturing conversation summaries, injecting short-term memory (default: 2 days), and providing keyword-based retrieval tools, it addresses memory loss caused by the model's limited context window. Implemented in pure JS with no external library dependencies, the plugin supports actively retrieving historical information when conversations exceed the context window and can share a data directory with opencode's memsearch.
Read Moredsh-memory-spaces: User-Autonomous Local Memory Spaces for DeepSeek Harness
`dsh-memory-spaces` is a standalone plugin for DeepSeek Harness (DSH) designed to address the difficulty of reusing decisions, constraints, and preferences across Sessions, which can be error-prone. It provides persistent, private "memory spaces" through a local SQLite database, allowing users to explicitly control which Sessions act as "sources" contributing content and which act as "consumers" reading content. Its core features include human governance, provenance, version rollback, and a pre-send review mechanism. The plugin does not rely on a global pool; all data remains fully local and requires users to manually establish connections, making it suitable for developers who need to reuse specific context across multiple conversations while prioritizing security and controllability.
Read Moredsh-mnemosyne: A Local Memory Plugin for DeepSeek Harness
`dsh-mnemosyne` is the local memory plugin for DeepSeek Harness, designed to address cross-session context consistency issues for agents. It encapsulates Mnemosyne’s memory logic as native DSH tools, with data stored in a local SQLite database. The plugin provides six tools—including memory storage, retrieval, and forgetting—as well as dashboard-based management, supporting automatic merging, filtering, and session isolation. Users can install and configure it via the CLI and leverage automatic injection and prefetching mechanisms to enable long-term memory without relying on cloud services.
Read Moredsh-hindsight-local-optimize: Local Memory Daemon Optimization
This plugin acts as a sidecar patch, taking over the startup and shutdown of the official `@vectorize-io/hindsight-coding-agents` daemon to resolve desktop-side issues such as cold-start latency, terminal black-window interference, and path restrictions. It provides core features such as a one-click toggle, automatic start/stop with DSH, hiding the terminal black window, and custom installation paths, improving the usability and stability of local memory on the desktop.
Read MoreXMemo: A Local-First Memory Plugin for DeepSeek Harness
XMemo is a native Cordis plugin that provides local-first and cloud-based memory capabilities for DeepSeek Harness. It communicates directly with the MemoryOS REST API, moving away from the traditional MCP bridge pattern, and offers features such as hybrid storage, an offline elastic queue, focused recall, and soft deletion. The plugin supports both OAuth 2.1 + PKCE and static API key authentication, and provides 16 tools to support work continuity. Developers can install it via npm or a local path and configure it in dsh to build persistent agents.
Read Moredsh-memory-graph: Local Knowledge Graph and Long-Term Memory Plugin for DSH
`dsh-memory-graph` is a localized, non-intrusive plugin designed for DeepSeek Harness that provides long-term memory and temporal knowledge graph capabilities. The SQLite-based plugin supports Chinese retrieval, hybrid relevance recall, and per-turn control, helping keep the graph maintainable and efficient across long conversations. Installation is simple and can be enabled by linking a local checkout, while the DSH Web interface supports visual inspection and management. Uninstallation only removes configuration and does not delete local databases or backup files, making it suitable for developers who need auditable, maintainable persistent storage.
Read Moredsh-memory-eternal: A Local Memory Plugin for the DeepSeek Harness
dsh-memory-eternal is a local memory plugin for the DeepSeek Harness, designed to address the persistence of session conclusions and on-demand recall of historical context. Maintained by EternalNight996 and released under the MIT license, the plugin consolidates conversations into a local Markdown Vault, supports Git-based management, and allows multiple Agents to share a common memory repository. Its core features include automatic knowledge card persistence, a memory_recall tool for on-demand retrieval, a built-in graphical knowledge base and knowledge graph, as well as a review center and recycle bin mechanism. The plugin supports integration with DSH, Claude Code, Codex CLI, and Cursor, and offers a standalone browser interface along with self-managed service options. Installation requires Node >= 22,
Read MoreDSH-AgentMemory: A Local Memory Plugin for Integrating AgentMemory into DSH
dsh-agentmemory is a DSH memory plugin (MIT licensed) maintained by elementor-i, designed to integrate a locally running agentmemory REST service into the DSH framework. This plugin does not require an MCP bridge and directly depends on a local service (default localhost:3111), providing agents with a stable memory backend for reusing session records and injecting recalled context. Core capabilities include providing 54 `memory_*` tools (including 8 core tools) and a `memory_http` escape hatch; automatically capturing sessions, prompts, tool calls, and other activities in the background; and supporting optional context injection. The default configuration emphasizes security, including credential redaction, timeout control, and non-blocking observation, with destructive tools disabled by default. Installation can be performed via
Read Moredsh-personalize: DSH Web Native Personalization Plugin
dsh-personalize is a DSH web plugin maintained by Zephyr-vibe, designed to address the issue of repeatedly configuring personalized settings across multiple sessions. The plugin is licensed under the MIT License, requires Node >= 20, and runs through the official public service without modifying the DSH core. Its core features include: 1. Custom instructions that inject users' long-term preferences into system prompts; 2. Local long-term memory, supporting manual addition or automatic collection by agents through tools like memory_create, with data persistently stored locally in the {DSH_HOME}/personalize/ directory, with a default limit of 50 entries; 3. Reply tone presets, offering various style options such as gentle and efficient. The plugin provides a bilingual interface in Chinese and English, and integrates automatically into the "Settings" menu after installation. Users can access it via commands or
Read More