AI Agent Hub
Back to plugins
🧠

dsh-local-vector-memory

Memory Updated 2026.08.31

Run the following command in DeepSeek Harness:

dsh plugin install liangxiaobing520/dsh-local-vector-memory

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

Install this local vector memory plugin in DeepSeek Harness by running dsh plugin install liangxiaobing520/dsh-local-vector-memory, source code at https://github.com/liangxiaobing520/dsh-local-vector-memory

About this plugin

DeepSeek Harness users often restate preferences, conventions, or caveats across sessions, only to lose them once the conversation ends. dsh-local-vector-memory puts long-term memory on your own machine: local embeddings via any OpenAI-compatible service, stored in a single SQLite file—no external vector database required.

Memories are written through three paths: manual memory_add, millisecond-level auto-capture when a user message hits recall cues, and session-flush extraction by DeepSeek v4-pro from the full conversation. Before every agent inference, the top-K relevant memories are injected into context with in-session LRU dedup. On write, a conflict check (similarity ≥ 0.86) nudges you toward memory_update instead of adding a duplicate. Deletion defaults to a soft-delete trash bin with one-click restore, and online backups use SQLite VACUUM INTO for a consistent snapshot, keeping the latest five by default.

The entire data path stays local: vector store, embedding inference, and keyword fallback all run on your machine. Only the end-of-session extraction call touches the DeepSeek official API. This plugin suits users who want full ownership of conversational memory, already run a local embedding service (e.g. Qwen3-Embedding 0.6B + llama-server), and work with the DSH web profile.

Use Cases

  • Auto-capture preferences and conventions across multi-turn conversations and recall them later
  • Extract key points from the full conversation at session end to reduce repetition
  • Keep memory local with SQLite; fall back to keyword matching when embedding is down

Best For

  • DSH users who already run a local embedding service (e.g. Qwen3-Embedding + llama-server)
  • Developers who want full ownership of conversational memory on their own machine
  • DSH web-profile users who prioritize data sovereignty and local-first workflows