topprismdata Mem0 Dynamic Graph Memory (Entity-Relation Extraction)
Context
Static OKF stores predefined edges. Mem0/Letta (2025) extract entities + relations dynamically from text. agy suggested upgrading our existing EntityRelationExtractor as P2: enables automatic concept discovery instead of manual OKF curation.
The core insight: knowledge graphs are useful only if they reflect real relationships, and those relationships must be extracted, not hardcoded.
Guidance
Extract from Paper/Discussion
from framework.src.memory.entity_extraction import EntityRelationExtractor
ext = EntityRelationExtractor()
text = """
LightGBM uses histogram-based gradient boosting.
XGBoost improves LightGBM with second-order gradients.
CatBoost handles categorical features natively, unlike LightGBM.
We evaluated F1 score on Spaceship Titanic.
Walk-forward validation prevents data leakage in time series.
"""
entities, relations = ext.extract(text)
# → 8 entities (techniques, metrics, competitions, concepts)
# → relations: "XGBoost improves LightGBM", etc.
for e in entities:
print(f"{e.type}: {e.name} ({e.mentions} mentions)")
for r in relations:
print(f"{r.source} --{r.relation}--> {r.target}")
With LLM Enhancement
def my_llm(prompt: str) -> str:
return openai_client.chat.completions.create(...).choices[0].message.content
ext = EntityRelationExtractor(llm_call=my_llm)
entities, relations = ext.extract(text)
# → 规则抽取 + LLM 抽取合并去重
Integrate with OKF Index
# 已存在 OKFIndex,可叠加 entity edges
from framework.src.memory.okf_index import OKFIndex
from framework.src.memory.entity_extraction import EntityRelationExtractor
okf = OKFIndex(okf_dir="docs/ml-agent-memory")
okf.build()
ext = EntityRelationExtractor()
entities, relations = ext.extract(paper_abstract)
# 把 entity edges 添加到 OKF 图
for e in entities:
okf.items[f"entity:{e.name.lower()}"] = MemoryItem(
id=f"entity:{e.name.lower()}",
content=f"{e.type}: {e.name}",
type="entity",
importance=0.6,
)
Why This Matters
| Static OKF | Dynamic Entity-Relation |
|---|---|
| Edges hardcoded in markdown | Edges extracted from text |
| Stale (doesn't reflect new content) | Always current |
| Misses implicit connections | Finds "X improves Y" relations |
| Manual curation required | Auto-scaling |
agy: "Mem0 dynamic graph memory enables cross-Competition knowledge transfer at scale."
When to Apply
When to Use
- Ingesting new paper or discussion
- Building knowledge graph from corpus
- Analyzing relationships between techniques
- When OKF misses connections you see in papers
When NOT to Use
- Short texts (LLM overhead not worth it)
- Highly specialized domain (entity patterns need extension)
- Real-time scenarios (extraction takes seconds)
Notes
- Pattern-based fallback: works without LLM, slightly lower recall
- Type coverage: technique / tool / metric / competition / concept
- Dedup is critical: same entity mentioned multiple times = single node
- Confidence scoring: relation extraction includes evidence snippet
- See also:
memory-hierarchy-management,time-series-walk-forward-validation
References
- Implementation:
framework/src/memory/entity_extraction.py - Inspired by: Mem0 (2024), Letta/MemGPT (2023), GLiNER (NER model)
- Pairing:
framework/src/memory/okf_index.py(existing static OKF)