Condenser Pipeline¶

The CondenserPipeline is a 4-stage memory compression system that reduces conversation history size while preserving essential information. It works alongside Sleep Mode and the Memory Synthesizer.

Stages¶

graph LR
    A[Full History] --> B[Observation Masking]
    B --> C[Amortized Forgetting]
    C --> D[Summarizing]
    D --> E[Windowing]
    E --> F[Compressed History]

1. Observation Masking¶

Removes redundant tool outputs that have already been processed. For example, if a file was read and its contents incorporated into the conversation, the raw file content can be masked.

2. Amortized Forgetting¶

Progressively reduces detail in older messages. Recent messages retain full detail; older messages are condensed proportionally to their age.

3. Summarizing¶

Groups related messages and replaces them with summaries that capture key facts, decisions, and outcomes.

4. Windowing¶

Applies a sliding window that keeps the most recent N messages at full fidelity, ensuring the model always has immediate context.

When It Runs¶

The condenser pipeline is triggered by the CompactionManager when:

• Context usage exceeds the configured threshold (default 80%)
• Sleep mode activates via MemoryConsolidator
• Manually via compaction scheduling

Relationship to Other Systems¶