ACE — Autonomous Cognitive Engine

ACE is an open-source cognitive runtime for building AI agents that can execute long-horizon tasks while maintaining explicit, persistent cognitive state.

Most LLM agents rely heavily on conversation history and lose context when a long task becomes complex or execution is interrupted. ACE treats the agent’s working state as a first-class object: goals, research plans, completed tasks, pending tasks, findings, sources, decisions, and unresolved questions can be tracked and checkpointed throughout execution.

What ACE does

ACE combines Gemma 4, LangGraph, MCP tools, and persistent state to create a research agent capable of:

• Breaking complex objectives into research tasks
• Using tools and MCP capabilities to gather evidence
• Coordinating multiple research agents
• Maintaining structured cognitive state throughout execution
• Checkpointing progress during long-running tasks
• Recovering from interruption and resuming from the previous state
• Continuing unfinished work instead of restarting from scratch
• Synthesizing accumulated evidence into a final report
• Providing full execution visibility through LangSmith tracing

The key idea

ACE is not simply another chatbot or wrapper around an LLM.

The core idea is to separate reasoning from cognitive state and execution.

Gemma 4 provides the intelligence for reasoning, planning, tool selection, and synthesis, while ACE provides the runtime that manages state, execution, checkpoints, recovery, and long-horizon task progression.

Example

A complex question such as:

“Should India invest heavily in domestic semiconductor manufacturing over the next 10 years?”

can be decomposed into technology, economics, supply chain, geopolitics, environmental impact, and international comparisons.

ACE can research several of these dimensions, checkpoint its state, be interrupted, restore its previous cognitive state, and continue with the remaining tasks before producing a final recommendation.

Architecture

                    USER OBJECTIVE
                          │
                          ▼
                ┌───────────────────┐
                │    ACE Runtime    │
                │                   │
                │ Goal / Plan       │
                │ Tasks / Progress  │
                │ Findings / State  │
                └─────────┬─────────┘
                          │
                          ▼
                     ┌─────────┐
                     │ Gemma 4 │
                     └────┬────┘
                          │
                 ┌────────┼────────┐
                 ▼        ▼        ▼
             Supervisor  MCP     Tools
                 │
                 ▼
             Researchers
                 │
                 ▼
          Cognitive State
                 │
          ┌──────┴──────┐
          ▼             ▼
      Checkpoint      Replan
          │
          ▼
        Resume
          │
          ▼
       Final Report