evo-memory

evo

Evolving memory for agents — recall before the model thinks, reflect after it answers.

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/

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:sqlite is 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