Project archive / 04
Memory-Aware Coding Agent CLI
mitosis-cli
A React and Ink coding CLI with branching ReAct, governed tools, Mempedia long-term memory, reusable skills, and sandboxed local execution.
- Archive
- 04
- Category
- Memory-Aware Coding Agent CLI
- Result
- 5 governed tools · 4 memory layers
System / highlights
What I delivered.
- 01
Built a functional branching ReAct loop that can continue linearly, fork viable strategies into child branches, and synthesize completed branches into one answer.
- 02
Reduced the planner surface to five governed tools—read, search, edit, bash, and web—with workspace boundaries and command-level sandbox checks.
- 03
Connected the agent to Mempedia over NDJSON and added an asynchronous classifier that saves qualifying knowledge across four memory layers after each turn.
- 04
Added local and remote SKILL.md discovery, activation, and injection, plus an integrated UI bridge for dialogue, traces, and memory inspection.
Archive / full note
Project notes.
A memory-aware coding CLI that combines branching ReAct, reusable skills, and governed local execution.
Overview
Mitosis CLI is a React and Ink coding agent that uses Mempedia for context and long-term knowledge management. It works with OpenAI and OpenAI-compatible model endpoints and keeps runtime state inside the current workspace.
Branching ReAct
The agent treats ReAct as a functional loop. A step may continue with one tool plan, fork into multiple child branches when strategies materially differ, or finish with a final answer. Completed branches are merged by a synthesizer, while depth, width, step budgets, and concurrency remain configurable.
Governed execution and memory
The planner sees only five top-level tools: read, search, edit, bash, and web. Workspace access is bounded, and shell commands run inside a project-local sandbox with checks for repository synchronization, remote transfer, privilege escalation, and download-then-execute patterns.
After a completed turn, an independent classifier evaluates the interaction against Mempedia’s four memory layers and queues qualifying saves asynchronously. Local SKILL.md files can be matched automatically or activated explicitly, while remote skill discovery supports reusable workflows without permanently expanding the base prompt.
Index / methods
Technology & methods.
- 01React
- 02Ink
- 03Branching ReAct
- 04Mempedia
- 05OpenAI-compatible APIs
- 06SKILL.md
- 07NDJSON
- 08Local Sandbox