dsh-memory-plugin: A DSH plugin that gives DeepSeek Harness cross-project long-term memory
x-000001/dsh-memory-plugin
A DSH plugin that gives DeepSeek Harness cross-project long-term memory with session, project, global and habit stores, progressive recall, and a zero-LLM core.
dsh-memory-plugin is a long-term memory extension for DeepSeek Harness 0.2.x that splits memory into four physically separated stores: session, project, global and habit, so information survives the end of a conversation and preferences carry across projects. It controls token use with a resident index plus on-demand expansion, and keeps the injected section byte-stable to reuse prompt caches. Habit learning follows an auditable pipeline that requires two human confirmations before a habit takes effect automatically. The whole core path runs without model calls, so writing, recall and forgetting work even with no model route configured.
How to Install
dsh plugin --profile web add github:x-000001/dsh-memory-plugin- Category
- Data & Knowledge
- Platform
- DSH Plugin
- Author
- x-000001
- Distribution
- Plugin
dsh-memory-plugin Key Features
dsh-memory-plugin Repo Summary
What Does It Do?
It is a DSH plugin for DeepSeek Harness that adds cross-project long-term memory to DSH 0.2.x as a Cordis plugin bundle. It solves the problem of context being lost when a session ends and preferences failing to carry across projects by keeping four physically separated stores: session, project, global and habit. Its core capabilities cover layered storage, token-efficient recall through progressive disclosure, and an auditable user-habit learning pipeline. The whole core path runs with zero LLM calls, so writing, recall, injection, forgetting and health reporting still work when no model route is configured.
Core Features
- Four separated stores: session, project, global and habit memory live in distinct roots; project memory travels with the repository, global preferences follow the DSH home directory, and the roots are validated as strictly non-nested.
- Progressive recall: a hard-capped resident index plus on-demand memory_recall and memory_expand, with a per-session cumulative budget and an observable degradation ladder.
- Cache-friendly injection: the injected section stays byte-identical while its content signature is unchanged, and the index only updates when memory content actually changes.
- Habit learning pipeline: a deterministic path from observation to threshold, optional judge, candidate zone, and two human gates, so a model can never promote a habit on its own.
- Zero-LLM mode: write, recall, inject, forget, consolidate and health reporting all work without a model route; the optional judge model defaults to off and fails open.
How to Use This Plugin?
After enabling this DSH plugin in DeepSeek Harness, everyday conversations accumulate memory automatically with no extra steps. To retrieve older information, call the recall tools to read the resident index first and expand specific entries on demand; project memory moves with the repository, so it survives a machine change. Habit entries never take effect automatically: they require two human confirmations on the host UI route, and no tool parameter can confirm or promote a habit. For a fully offline setup, leave the judge model off and the entire memory path keeps working.
How to Troubleshoot This Plugin?
If startup reports conflicting memory roots, the project, global or habit directories alias or nest each other; follow the official instructions to make the four storage roots independent and non-overlapping. If a configured judge model fails, the pipeline fails open and continues, so turn the judge off first to check whether the main path is healthy.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: x-000001/dsh-memory-plugin. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.
DSH Plugin FAQ
