AI Agent Hub
Back to plugins
🧠

dsh-long-memory

Memory Updated 2026.08.26

Run the following command in DeepSeek Harness:

dsh plugin install wwskills/dsh-long-memory

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install wwskills/dsh-long-memory in the DeepSeek Harness terminal to install; the source is available at https://github.com/wwskills/dsh-long-memory

About this plugin

Every time a DeepSeek Harness session ends, the context vanishes with it-corrections you made last round, preferences you built up, rules you encoded, all of it gone by the next conversation. dsh-long-memory builds a SQLite-backed persistent memory layer so the agent remembers what matters. Out of the box it ships with FTS5 full-text search (native CJK support, zero config), optional embedding via a local Ollama service or any OpenAI-compatible API, and a BM25 plus vector RRF hybrid recall pipeline that surfaces the right memories in any new session.

Beyond recall, the plugin closes a self-evolving learning loop. When the user says a signal word-such as "wrong", "should be", or the Chinese equivalents (15 configurable patterns across 8 CN and 7 EN)-or when a tool call or agent step fails, the correction is captured in real time. An LLM then distills it into a structured lesson that enters a full rule lifecycle: proposed, approved, rejected, archived, or promoted to AGENTS.md. Approved rules are auto-injected into agent context within an 800-token budget, with hit-count tracking, Jaccard-overlap conflict detection, and automatic archival after 90 days of inactivity.

On the tooling side, eight mem_* tools let the agent read, write, search, and scope memories directly. A complete REST API and a four-tab browser panel (Corrections, Rules, Memories, Persona) give you full visibility, while every operation lands in an append-only audit log. The plugin is built for DeepSeek Harness users who need cross-session persistence, want their agent to learn from user feedback, or require a defensible audit trail. Requires Node 22.5 or later and DeepSeek Harness 0.1.0-rc.2 or above.

Use Cases

  • Retain user preferences and corrections across sessions so the agent does not start from scratch each round
  • Auto-extract rules from user feedback and tool errors, then inject approved rules into agent context for continuous improvement
  • Review corrections, rules, memories, and persona in a four-tab browser panel, backed by an append-only audit log for compliance

Best For

  • Harness users who need cross-session memory persistence and cannot afford to lose historical context
  • Developers and teams who want the agent to learn automatically from user corrections and accumulate rules
  • Production operators who require auditable operations and traceable memory management