AutoDream & Compaction

Relevant source files

The following files were used as context for generating this wiki page:

• README.md
• internal/engine/workflow/plan.go
• internal/engine/workflow/prompts/base.md
• internal/engine/workflow/prompts/autonomous.md
• pkg/autodream/autodream.go (Referenced via README.md architecture and package summary)
• pkg/compaction/compaction.go (Referenced via README.md architecture and package summary)

AutoDream and Compaction are critical context management systems within the M31 Autonomous (M31A) engine. They ensure that long-running coding sessions remain within the Large Language Model's (LLM) context window constraints by consolidating history and summarizing previous interactions. While standard coding assistants often lose context as conversations grow, M31A uses these systems to "own the loop" and maintain a coherent mental model of the task across hundreds of tool calls.

Sources: README.md:27-40, internal/engine/workflow/prompts/base.md:12-25

AutoDream Context Consolidation

AutoDream is a specialized system designed to compress long conversations automatically when the context window nears its limit. It prevents context overflow by merging redundant information while preserving essential data required for task execution.

Protection Logic

AutoDream does not compress all messages equally. To ensure the agent does not lose its "identity" or current progress, specific message types are protected:

• The initial user goal.
• Active tool calls and their results.
• The implementation plan.
• The last 5 messages in the current buffer.

Sources: README.md:173-177, internal/engine/workflow/prompts/autonomous.md:35-45

Workflow Integration

The consolidation process is triggered automatically on context window overflow. It can also be manually invoked via the /compress command in the terminal UI.

The diagram above illustrates how AutoDream transitions from a trigger event to a condensed state without losing critical task data.
Sources: README.md:173-177, README.md:210-220

Session Compaction

Compaction operates at the package level (pkg/compaction) to provide automatic session summarization. Unlike the immediate consolidation of AutoDream, Compaction focuses on replacing detailed historical logs with LLM-generated summaries.

Key Functionalities

• Summarization: Collapses multiple turns of conversation into a concise narrative.
• Replacement: Swaps out original verbose messages for summarized versions in the session persistence layer.
• Token Efficiency: Reduces the total token count of the session history while maintaining the logical flow of decisions made.

Sources: README.md:315-325, internal/engine/workflow/plan.go:50-65

Comparison of Management Systems

M31A employs multiple layers of memory management to balance cost, performance, and accuracy.

System	Purpose	Persistence	Trigger
AutoDream	Real-time context compression	Volatile/Session-based	Context window overflow (>95%)
Compaction	Historical summarization	Saved to STATE.md/Session JSON	Post-phase transitions
Learning Ledger	Cross-session knowledge	Permanent Markdown-backed	Session completion (Ship phase)

Sources: README.md:144-160, README.md:315-330, internal/engine/workflow/plan.go:150-165

Technical Implementation and Safety

The implementation of these systems includes reentrancy guards and safety heuristics to prevent the agent from entering a loop of continuous compression.

Context Awareness Rules

The engine enforces specific "Read-Before-Write" rules that interact with consolidated context. Even after a compaction event, the agent is instructed to:

1. Map the dependency graph using CodeMap or Grep.
2. Read the full content of relevant files (target, imports, callers).
3. Verify project conventions before applying summarized knowledge.

Sources: internal/engine/workflow/prompts/base.md:33-50, internal/engine/workflow/prompts/context-awareness.md:10-30

Token Calibration

M31A uses tiktoken for estimation combined with Exponential Moving Average (EMA) self-calibration. This allows the AutoDream system to predict exactly when a compaction is required.

• Warning Banner: Triggers at 80% context usage.
• Hard Reject/Auto-Compress: Triggers at 95% context usage.

Sources: README.md:178-182, internal/engine/workflow/plan.go:80-95

Conclusion

AutoDream and Compaction are foundational to M31 Autonomous's ability to handle complex, multi-file software engineering tasks. By intelligently summarizing history and protecting critical goal-oriented data, these systems allow the agent to operate efficiently within LLM token limits while maintaining high code quality and architectural consistency across extended sessions.

Sources: README.md:12-25, internal/engine/workflow/prompts/autonomous.md:10-25