mem0-dynamic-graph-memory

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,
    )