Introduction

The core concept of DeepSeek Harness (DSH) is “everything is a plugin.” In mathematical learning scenarios, notation conventions, theoretical preferences, unfinished proofs, techniques, and counterexamples all require long-term accumulation and continuous refinement to form a coherent system. General-purpose chat AI often treats each conversation as an isolated Q&A and cannot carry over historical context.

dsh-math-memory is a DSH assistant for mathematics notes, residing in Obsidian’s right sidebar. It provides the agent with cross-session hierarchical memory, a unified retrieval protocol, and a maintenance strategy for memos, ensuring that each new session continues from the previous state.

Plugin Overview

  • Name: dsh-math-memory
  • Maintainer: maple110011
  • Category: Memory
  • License: MIT
  • Positioning: A DeepSeek Harness assistant for mathematics notes, residing in Obsidian’s right sidebar.

Core Features

Hierarchical Memory System

The plugin constructs a five-layer memory structure:
1. Profile: Semantic layer, used to define context.
2. Topic: Navigation layer, used for quick localization.
3. Record: Atomic card layer, with hook: retrieval blocks and verification levels (✅⚖️❓).
4. Raw Evidence: Append-only records, preserving raw conversation evidence.
5. Inbox: Memo layer, used to capture ideas, which can later be refined into formal records.

Unified Retrieval and Verification

  • Unified Search: Uses the note_recall protocol to retrieve notes and all memory layers (hook cards, memos, topics, theorem/evidence indexes) in a single BM25 ranking.
  • Read Verification Protocol: The system distills the query into a challenge and reads the first 2-3 hits for human judgment. If the results are empty or weak, it reformulates the query once, then acknowledges “not in the vault” rather than fabricating content.

Symbolic System and Maintenance

  • Symbolic System Maintenance: Maintains memory/notation.md, including tables for adopted/candidate/rejected items and a revision history to prevent notation drift.
  • Daily Audit: Automatically scans strong/weak/unused/duplicate candidates/unverified cards and checks structural integrity (missing sources/broken links/missing index lines).
  • Memo Reminders: Provides refinement reminders for outdated memos (inbox > 7 days, refinement > 3 days) or memos relevant to the current session.

Cross-Session Context and Response Protocol

  • Cross-Session Context: Distills historical DSH sessions (zstd JSONL) into bounded Q&A prompts for use in the current session.
  • Response Quality Protocol: Gives intuition before formalization, anchors new material to existing notes, performs Socratic correction, and checks for low-frequency issues.

Control Panel and Feedback

  • Obsidian Memory Panel: Displays a status bar (status of each layer + last audit time), a list of cards requiring review (supports ✅/❌/stale/archive actions), an evidence timeline, and a Chinese audit summary.
  • Feedback Loop: Replies include Memory basis: <card title> — [✅ Correct] [❌ This card is incorrect] links, correcting memory cards through the backend /feedback endpoint.

Independence

By default, it does not load @linxin666 UI plugins (such as Skin Center, Task Board) to avoid dependency conflicts. Skin Center can be manually enabled in settings.

Installation and Activation

Use the official CLI to install:

dsh-math-memory install

Prerequisites:
* DeepSeek Harness >= 0.1.7-rc.2
* Node.js >= 22.5

Typical Usage

  1. Search and Read: Use note_recall to search notes and memory, and judge signal strength based on coverage metrics.
  2. Strategy Handling: Use note_strategy to handle proof or construction problems.
  3. Create Notes: Use note_create to create a new note (overwriting is rejected).
  4. Correct Memory: Click the feedback link in a reply to correct memory cards.
  5. Review and Archive: Review cards and archive them in the memory panel.

Applicable Scenarios and Notes

  • Applicable Scenarios: Users who need to maintain a mathematics knowledge system over the long term, want AI to remember contextual details, and are accustomed to using Obsidian for note management.
  • Version Notes: The current version (0.7.x) is a prototype and lacks an active teaching loop and invocation system; the 1.0 version will include complete teaching and invocation features.
  • Dependency Notes: UI plugins are not loaded by default; Skin Center requires additional installation if needed. The note_search feature is deprecated.
  • Ecosystem Notes: This plugin follows the DSH “everything is a plugin” ecosystem philosophy. Community directories (such as SkillHub) have no official affiliation or hierarchical relationship with DeepSeek. Please inspect the source code and license before installation.

Summary

dsh-math-memory solves the cross-session context discontinuity problem in mathematics notes through hierarchical memory and unified retrieval. It provides an auditable and correctable memory foundation, suitable for technical users who need to deeply consolidate knowledge systems. Project address: GitHub | Community Directory.