Introduction

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin,” covering different needs in agent development through extensible capabilities. When developing coding agents, a common problem is context loss: the agent cannot remember previous decisions, lessons, or project structure across sessions, leading to reinventing the wheel or repeating mistakes.

Veyra aims to solve this problem. It provides DSH with a Dual-Brain architecture for “persistent engineering memory” and “optional AST code intelligence.” The plugin does not depend on external API keys and runs as pure ESM, giving agents cross-session memory capabilities.

What Is Veyra

Veyra is a DSH plugin maintained by LoveDoLove. Its core function is to provide persistent, evidence-consistent engineering memory and optional structured code analysis. It separates “engineering decision memory” (such as root causes and fixes) from “code structure truth” (such as call graphs and symbol tables), helping agents both leverage past experience and accurately locate code structure when executing tasks.

Installation and Enablement

Installation relies on DSH’s plugin management tool. After installation, restart or reload the configuration file to activate the plugin.

  1. Run the installation command:
dsh plugin --profile web add @lovedolove/veyra
  1. Restart DeepSeek Harness or reload the current configuration file; Veyra will activate automatically.

Prerequisites: Node.js must be installed, with a version in the range >=22.5.0 and <25.0.0.

Dual-Brain Architecture

Veyra answers different types of questions through two complementary layers:

  • Veyra Memory (Memory Brain): Stores engineering-level lessons learned, including root causes, decision constraints, causal chains (symptom → root cause → fix → verification result), and verification status. This is the accumulated “knowledge state” across sessions.
  • Code Intelligence (Code Brain): Provides AST-level structured facts, including symbols, files, call graphs, and class hierarchies. This data always comes directly from the current repository, not from memory.

These two layers do not conflict. When code structure changes, the code intelligence layer marks related memories as “possibly stale,” but it never silently overwrites the contents in the memory brain.

Core Features

  • Persistent Engineering Memory: Each memory carries evidence anchors, verification status (unverified → reviewed → verified), and authority level. Veyra does not modify repository files; repository files remain the authoritative source.
  • Hybrid Search: Supports FTS5 BM25 full-text search, symbol/path weighting, semantic overlap, and intent affinity (why/how/what happened) with composite scoring, along with a transparent score breakdown.
  • Change Awareness: Agents automatically detect repository changes. Change summaries guide memory retrieval; Veyra checks whether memory evidence anchors point to modified or deleted code and applies retrieval penalties to stale records.
  • Health Checks and Maintenance: Provides the veyra_health command and a Web interface to view memory health. The built-in memory-maintenance skill provides lifecycle management actions (retain, override, invalidate, forget, tombstone, revalidate) that require explicit execution requests.
  • Web Interface: Access /veyra in DSH Web to view the read-only “knowledge observatory” and the interactive “network graph.”

Code Intelligence (Optional)

AST-level code intelligence is optional. It relies on the external tool codebase-memory-mcp to provide structured data.

  1. Install the native engine (if enabling Code Intelligence):
curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash
  1. Degraded mode: If the binary is not installed, Veyra can still run its memory features normally. All Code Intelligence tools will return structured errors containing installation guidance, without interrupting the memory features.

Typical Usage

Veyra provides a series of Slash Commands for interaction between agents and users.

  • View recent records:
    /veyra recent
  • Recall memory:
    /veyra recall sqlite wal
  • Deep inspection and verification:
    /veyra inspect vey_...
  • Memory management:
    /veyra forget vey_...      # soft forget, preserving inspectability
    /veyra promote vey_... canonical # explicitly promote to a canonical memory

In the Web interface, access /veyra to view the knowledge graph and local network graph. Access /veyra?id=<record> to focus on the one-hop local graph for a single record.

Notes

  • Package Size: The plugin is a pure ESM package and is approximately 180 kB after packaging.
  • Permissions: Veyra runs with the permissions of the current DSH process and reads repository contents.
  • Security: Check the source code and license (MIT) before use.
  • Repository Protection: Veyra only reads the repository and never writes files. The authority of memories comes from the repository files themselves.
  • No External Dependencies: Basic memory features do not require API keys or external services.