A coding agent harness that runs on 28 MB and never forgets

What it does

jcode is a Rust-based terminal harness for AI coding agents. It wraps LLM providers into a keyboard-driven TUI that supports multi-session workflows, side panels for file diffs and live views, and inline Mermaid diagrams. The README emphasizes resource efficiency and an agentic memory system over raw model capabilities.

The interesting bit

The memory system embeds every conversational turn as a semantic vector, then passively surfaces relevant past context via cosine similarity without burning tokens on explicit retrieval calls. A background memory sideagent periodically extracts, verifies, and consolidates memories—including staleness and conflict checks—while the entire UI cold-boots in roughly 14 milliseconds.

Key highlights

• Baseline RAM footprint of ~28 MB per session (or ~168 MB with local embeddings enabled), scaling by roughly 10 MB per additional session—significantly leaner than the competitors listed in the README’s benchmarks.
• Cold-start latency of ~14 ms to first frame and ~49 ms to first input, measured against rivals that take hundreds or thousands of milliseconds.
• Built-in semantic memory graph with ambient consolidation, plus explicit memory tools and session search for traditional RAG.
• Custom Rust Mermaid renderer with no browser or TypeScript dependency, reportedly 1800× faster than typical alternatives.
• Cross-platform TUI with real-time side panels that can act as diff viewers or live file monitors.

Caveats

• The published benchmarks pit a dev build (v0.9.1888-dev) against stable releases of competing tools, so treat the comparison tables as directional rather than definitive.
• Several feature sections are truncated in the provided README, leaving the full scope of MCP integration and “infinite customizability” largely unexplained.

Verdict

Worth a look if you live in the terminal and want a fast, memory-efficient agent harness with persistent context across sessions. Skip it if you need a mature IDE integration or a fully vetted stable release today.