Skip to content

feat(architecture): tiered memory architecture (L1-L5) for retrieval cost reduction #605

Description

@itlackey

Problem

All memory passes query the full stash regardless of abstraction level. Quick-shredder passes only need high-level patterns but scan raw session extractions too.

Proposed change

Organize the AKM stash into 5 semantic levels following the TiMem pattern:

  • L1-L2: Raw session extractions and immediate facts
  • L3-L4: Distilled patterns, recurring insights, project-specific knowledge
  • L5: High-level user preferences and persistent behaviors

Quick-shredder passes query L3-L5 for triage relevance. Deep consolidation descends to L1-L2. Expected benefit: 40-60% reduction in per-pass retrieval token cost as most quick passes only need the high-abstraction layers.

Source

AKM Improve Pipeline Optimization Report 2026-06-12, Section 5J

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions