evo
已验证 · 实测可装 TIZ36
功能简介
agent memory plugin based on CLP
可用 — 实测通过,早期项目
agent memory plugin based on CLP 实测能干净安装、正常启动。早期项目,但功能可用。
「已验证」表示我们的自动化 CI 在干净 profile 里实际执行了 dsh plugin add 并启动成功——仅此而已。功能描述与版本兼容性均为作者声明。这不是安全审计,也不代表对第三方代码的背书。
README
evo is a provider-neutral memory service that gives AI coding agents durable, cross-session memory. It recalls relevant knowledge into the system prompt before a model step, and reflects completed turns into structured memory afterwards — so an agent keeps what it learned across sessions without you writing it down.
Why evo?
Local CLI agents (Claude Code, Codex, DeepSeek Harness) are stateless by default: each session starts from zero. User preferences, project conventions, pitfalls stepped into three times — all evaporate when the session ends.
evo solves this with a recall + reflect loop:
- Recall: Before each prompt, assemble relevant memories into the agent's context
- Reflect: After each successful turn, distill the conversation into structured memory items
The result: your agent becomes an experienced colleague who remembers the project.
Documentation
📖 Full documentation: https://evo-5.gitbook.io/evo/
| Section | Description |
|---|---|
| GitBook Docs | Primary documentation site |
| Getting Started | Requirements and installation |
| Host Integrations | DeepSeek Harness, Claude Code, Codex |
| Concepts | How recall + reflect works |
| Configuration | All settings and environment variables |
| API Reference | HTTP and service APIs |
| Development | Contributing guide |
Quick Start
DeepSeek Harness
./install_evo_dsps.sh
Claude Code
/plugin marketplace add TIZ36/evo
/plugin install evo
Codex
codex plugin marketplace add TIZ36/evo
codex plugin add evo@evo
Key Features
- Structured memory: Every item has a scope, kind (fact, constraint, procedure, skill…), tags, and source
- First-class skills: Procedural SOPs separate from declarative memories, with lessons learned tracking and a dedicated Skills tab in the Memory panel
- Batch distillation: Turns are collected and distilled in batches for better memory quality
- Local by default: SQLite storage via Node's built-in
node:sqlite— nothing leaves the machine except reflection model calls - Workspace import: Existing
CLAUDE.md,AGENTS.md,.codex/, and similar files are imported automatically - Provider-neutral: Works with DeepSeek Harness, Claude Code, and Codex through adapters
Status
evo is currently alpha (0.3.x). The core loop is stable and used daily, but APIs may change.
Current Limitations
- Recall uses deterministic SQLite filtering and ranking, without embeddings
- Consolidation runs automatically on the slow path (floor 24h / 72h converged, or large replay buffer); call
consolidate()or use the web panel for manual runs node:sqliteis still marked experimental by current Node releases (it ships with Node 22+)
Development
pnpm install
pnpm check # test + typecheck + build + rule scan
See Contributing for the full development guide.
Project Rule
This is a personal open-source project. Company identity and sensitive information must never appear in the repository. Team attribution is "Paper team" only. See Project Rules for details.
License
MIT © TIZ36
安装
装一次目录插件,之后本站所有插件都能让 DeepSeek Harness 自动找、自动装:
dsh plugin add dshbase-catalog 然后对 agent 说「帮我装 evo」,它会在目录里找到并自动安装。文档:dshbase-catalog · 已验证场景包。
该插件是 GitHub 源码(未发 npm)——直接从仓库装:
Web profile:
dsh plugin --profile web add github:TIZ36/evo Headless(CLI)profile:
dsh plugin --profile headless add github:TIZ36/evo 实测报告
验证通过:从 GitHub 源码完成 L1 安装 + L2 加载 + L3 运行(dsh 0.1.0-rc.6)。