[Feature]: Compressed Memory Snippets for Token Efficiency

[zircote]

Problem Statement

When memories are included in LLM context, they often contain verbose content that consumes unnecessary tokens. Mem0 reports 80% token cost reduction through intelligent compression while maintaining semantic meaning.

Currently Subcog includes full memory content in recalls, leading to:

• Higher token costs
• Reduced context window availability
• Potential truncation of important content

Proposed Solution

Implement intelligent memory compression:

1. Snippet Generation: Create concise summaries of memories
2. Adaptive Compression: Adjust verbosity based on available context
3. Key Point Extraction: Identify and preserve critical information
4. Expandable References: Link to full content when needed

Compression strategies:

• LLM-based summarization (highest quality, highest cost)
• Extractive summarization (fast, preserves original text)
• Structured extraction (facts, decisions, actions)
• Hybrid based on memory type

Proposed API:

pub struct CompressionConfig {
    strategy: CompressionStrategy,
    target_ratio: f32,          // e.g., 0.2 = 20% of original
    preserve_keywords: bool,
    include_source_ref: bool,
}

pub enum CompressionStrategy {
    None,
    Extractive { sentences: usize },
    Abstractive { model: String },
    Structured { template: String },
    Adaptive { token_budget: usize },
}

MCP tool parameters:

subcog_recall:
  query: "authentication patterns"
  compression: "adaptive"
  token_budget: 500  # Max tokens for all results

Response format:

{
  "memories": [
    {
      "id": "abc123",
      "snippet": "Auth: Use JWT with refresh tokens, 15min expiry",
      "full_content_ref": "subcog://memory/abc123",
      "compression_ratio": 0.15
    }
  ]
}

Alternatives Considered

• Full content always (current) - token inefficient
• Truncation only - loses end of memories
• User-provided summaries - maintenance burden

Additional Context

• Mem0 claims 80% token reduction
• Extends existing detail levels (light/medium/everything)
• Important for cost-conscious deployments

Breaking Change: No
Priority: Nice to have
Contribution: Yes, with guidance