AI Agent Hub
Back to plugins
🧠

dsh-long-term-memory

Memory Updated 2026.08.26

Run the following command in DeepSeek Harness:

dsh plugin install windrover/dsh-long-term-memory

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

Run dsh plugin install windrover/dsh-long-term-memory in your DeepSeek Harness terminal to install the plugin; the source is hosted at https://github.com/windrover/dsh-long-term-memory .

About this plugin

After a long conversation ends, user profiles, project decisions, and cross-session preferences tend to vanish. The next new session starts with the model knowing nothing, and you end up repeating the same context. dsh-long-term-memory gives DeepSeek Harness a zero-dependency, layered long-term memory layer: no vector database, no extra model calls, deterministic BM25 keyword retrieval with CJK awareness, and automatic memory injection before every request assembly.

The core design centres on a three-tier store: user profile, global cross-project facts, and workspace-specific memories, each persisted in its own JSONL file. Retrieval and injection follow user, global, workspace priority. Writes go through threat-pattern scanning and an optional human-approval gate; recall is explainable and reproducible; a calibrated injection wrapper on systemPrompt.context() honestly declares provenance and uncertainty so the model can judge relevance on its own; memory_correct provides a correction loop that marks stale records as superseded and removes them from future injection.

Who it is for: developers iterating on the same project long-term who need the model to keep architectural decisions and constraints in mind; teams that prefer deterministic, auditable recall over an embedding black box; and anyone who wants a lightweight memory layer without deploying SQLite or a vector service.

Use Cases

  • Retain user profile, decisions, and constraints across sessions so context is never lost
  • Isolate memories per scope when juggling multiple projects without cross-contamination
  • Audit writes with threat scanning and the correction loop; stale records are marked superseded and excluded from injection

Best For

  • Long-term project developers who need the model to retain architectural decisions and constraints
  • Teams that value explainable, reproducible recall over an embedding black box
  • Lightweight users who want zero-setup memory without deploying SQLite or a vector service