Foreword

Local CLI agents (such as DeepSeek Harness, Claude Code, and Codex) are stateless by default. At the end of each session, contextual information such as user preferences, project conventions, and encountered pitfalls is lost. evo aims to provide these agents with a cross-session memory mechanism, enabling them to remember and reuse information from previous conversations.

What Is This

evo is a provider-neutral memory service designed to provide persistent, cross-session memory capabilities to AI programming agents. Through a “recall + reflection” loop, it allows agents to retrieve relevant knowledge before reasoning and consolidate structured memories after responding. The project is maintained by TIZ36 and is currently in the Alpha stage.

Core Features

evo provides the following core capabilities:

  • Structured Memory: Each memory item includes scope, type, tags, and source, supporting fine-grained management.
  • First-Class Skills: Programmatic SOPs (Standard Operating Procedures) and learned lessons are managed as independent skills, distinct from regular factual memories.
  • Batch Distillation: Conversation content is batch-converted into structured memory items to improve memory quality.
  • Local-First: Data is stored in SQLite by default, relying on the Node.js node:sqlite module. Data never leaves the local environment (except when invoking a reflection model).
  • Workspace Import: Supports automatically importing existing documents such as CLAUDE.md and AGENTS.md as initial memories.
  • Multi-Platform Adaptation: Supports DeepSeek Harness, Claude Code, and Codex through adapters.

Installation and Enablement

In the DeepSeek Harness environment, use the officially provided installation script. Run the following command:

./install_evo_dsps.sh

After installation is complete, the plugin will run as a plugin within the DSH process.

Typical Usage

The core workflow of evo consists of two phases:

  1. Recall: Before the model reasons or generates a response, evo retrieves relevant memories from the database and injects them into the agent’s context.
  2. Reflection: After the model successfully responds, evo distills the key points of the current conversation into structured memory items and saves them.

Applicable Scenarios and Notes

  • Version Status: The current version is Alpha (0.3.x). The API may change, and it is not recommended for critical production workflows.
  • Recall Mechanism: It currently uses deterministic SQLite filtering and ranking, without relying on vector embeddings. It is suitable for scenarios requiring high exact-match accuracy.
  • Memory Consolidation: Automatic consolidation typically runs on the slow path (with a minimum interval of 24 hours or 72-hour convergence, or when the replay buffer is large). If manual triggering is required, it can be executed through the consolidate() function or the Web panel.
  • Environment Requirements: Node.js 22 or later is required, and node:sqlite is still marked as an experimental feature in the current Node version.
  • Project Nature: This is a personal open-source project. Company identities and sensitive information are strictly prohibited from appearing in the repository.

Short Conclusion

Through structured memory management, evo resolves the issue of state loss in local agents, enabling them to maintain coherence across sessions like an experienced colleague. For more details, refer to the project directory page and the GitHub repository.