Skip to content

CEM Tree Memory: Branch-classified retrieval on BEAM (78%) and MABench (77.4% F1) #965

Description

@CEMTrading888

CEM Tree Architecture: Branch-Classified Memory Beats Flat Embeddings

We've been benchmarking memory retrieval systems on BEAM and MABench. Our tree architecture — where context is classified into hierarchical branches (Roots → Trunk → Sector → Leaf) and questions route to the correct branch — consistently outperforms flat embedding stores.

BEAM Results (DeepSeek V4 Pro)

Method Score
No retrieval (bare LLM) ~38%
Standard retrieval 77.2%
CEM Tree (branch-classified retrieval) ~78%

MABench Accurate Retrieval

  • CEM tree architecture: 77.4% F1
  • Published baselines: Claude 3.7 Sonnet 49.6%, HippoRAG-v2 41.6%

Why It Works

Flat memory stores retrieve by vector similarity — "what's semantically nearby." But memory questions aren't about nearby — they're about category. "What time was my meeting?" routes to Events. "What did we decide about the API?" routes to Technical. Branch classification before retrieval means the model sees relevant context instead of 10K tokens of noise.

The tree is model-agnostic. Any LLM plugged in gets smarter. No fine-tuning.

Full results: cem888.ai/benchmarks
creator@cem888.ai

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions