01 — Problem

The "Chatbot" Limitation

Standard LLM interfaces are reactive: they wait for a prompt, provide a text response, and stop. True productivity requires an agent — a system that can break a goal into sub-tasks, execute tools, observe the result, and iterate until the goal is achieved without human intervention at every step.

Cyclic LangGraph-based loop: Reason — Act — Observe — Repeat
Persistent Keyword-indexed markdown shards for cross-session recall
Grounded Persistent shell, vision OCR, and headless web-browsing
02 — Architecture

The Agentic Loop

Momobot is not a linear pipeline. It is a state-machine defined by a LangGraph, allowing the LLM to loop back to reasoning after every tool execution.

USER INPUT ENTRY REASONING NODE TOOL EXECUTION COMPACT Tool Observation Loop

Core Innovation: Unlike standard agents that might hallucinate a sequence of actions, Momobot uses a task_state.json to maintain a structured plan. It tracks pending → running → done statuses, making it resilient to crashes and capable of complex multi-step recovery.

03 — Tool Ecosystem

Beyond Text Generation

To be useful, an agent needs a "body." Momobot's toolset is designed for high-fidelity interaction with the local OS and the web.

Category Tool Capability
System bash_shell Persistent PowerShell session (cwd and env survive)
Files str_replace_tool Surgical, diff-based file editing (no accidental overwrites)
Vision view_image Playwright rendering — Vision LLM analysis
Knowledge memory_tools Shard-based long-term storage (Keyword index)
Planning init_task Graph-based dependency tracking for goals
04 — Context Engineering

The Compaction Problem

Long sessions eventually hit the LLM's context limit, leading to "forgetfulness" or performance degradation. Momobot implements an active Compaction Trigger. 

# Logic for context health maintenance
if current_tokens > COMPACTION_THRESHOLD:
    # Preserve most recent messages for immediate flow
    history_to_summarize = messages[:-RECENT_WINDOW]

    # Generate a dense state summary
    summary = llm.generate_summary(history_to_summarize)

    # Update rolling session state and prune history
    session_state.update(summary)
    messages = messages[-RECENT_WINDOW:]

This ensures that critical decisions made 50 turns ago are preserved in a condensed summary, while the immediate conversation remains fluid and high-resolution.

05 — Impact

From Chatbot to Co-Engineer

Momobot transforms the AI interaction from "asking for code" to "assigning a project." It is capable of:

  • Autonomous Refactoring: Reading a codebase, identifying a pattern, and applying surgical edits across multiple files.
  • Research Loops: Searching the web, fetching content, and synthesizing a report into the Obsidian vault.
  • Visual Debugging: Rendering a web page, analyzing the layout via vision, and fixing CSS issues.
  • Persistent Learning: Remembering a user's technical preferences across weeks of different sessions.
Technical Repository