AIToday
Large Language ModelsAI Coding AssistantsOpen-Source AIZenn AI/MLPublished: Nov 22, 2026, 16:00 JST

MIT-licensed MomongaCode tackles long-session AI coding agent drift

Send on LINE
MIT-licensed MomongaCode tackles long-session AI coding agent drift

3 Key Points

  1. What happened

    MomongaCode is out under the MIT license, a terminal AI coding agent that works with Ollama, LM Studio, and Apple fm locally, plus OpenAI and xAI subscriptions or Anthropic, OpenAI, and Open Responses API keys; the command is momonga.

  2. Why it matters

    It targets common failure modes where agents degrade in long sessions, growing context and token costs, and unreliable safety and completion checks — addressing these in the runtime instead of leaving them to the model.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

MomongaCode was built by a developer who uses terminal AI coding agents daily and noticed that they often break down in longer sessions. The agent addresses five specific problem areas: memory and context degradation, token consumption inflation, ambiguous safety and completion checks, fragility during long runs and updates, and inconsistent behavior when switching models. For memory, it stores entries with file paths and SHA-256 hashes in a database, automatically flagging stale entries when files change, and it requires human approval before new memories are activated. Token-saving features such as prompt cache stabilization and lazy tool loading are on by default, and users can benchmark the impact. Safety is enforced at the OS level with sandboxes and default network blocking, and Z3 SMT solver checks permission rules for gaps. The agent also uses TLA+ specifications for core state transitions and separates everyday responses from a polish mode that cross-checks against actual code. A comparison table positions it against seven other coding agents, highlighting its built-in approach to memory freshness, token optimization, sandboxing, and verification. The tool is installable with a single command and is positioned as suitable for both local and hosted models.

FAQ
What models can MomongaCode use?
It runs with local models via Ollama, LM Studio, and Apple fm, as well as subscriptions like OpenAI and xAI, or API keys from Anthropic, OpenAI, and Open Responses.
How does MomongaCode handle memory freshness?
It stores memories with file paths and SHA-256 hashes, automatically flagging them as possibly stale when the file changes, and requires human approval before new memories become active.
What safety features are built in?
It uses OS-level sandboxes (Seatbelt on macOS, bubblewrap on Linux), blocks network access by default, and includes Z3 SMT solver checks for permission rules and runtime verification of completion conditions.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleASIO's Mike Burgess flags AI data centre foreign-interference risk