AI Memory Session Graph

Model sessions, messages, entities, claims, tool calls, and evidence as graph records. This keeps memory inspectable and allows retrieval by both structure and properties.

from tonggraph import Graph

memory = Graph()

session = memory.add_node(
    "session:2026-06-22",
    labels=["Session"],
    properties={"user": "demo", "topic": "graph memory"},
)
message = memory.add_node(
    "message:1",
    labels=["Message"],
    properties={"role": "user", "text": "Remember that Alice likes graph RAG."},
)
alice = memory.add_node(
    "entity:alice",
    labels=["Entity", "Person"],
    properties={"name": "Alice"},
)
claim = memory.add_node(
    "claim:1",
    labels=["Claim"],
    properties={"text": "Alice likes graph RAG.", "confidence": 0.92},
)

memory.add_edge(session, message, "HAS_MESSAGE")
memory.add_edge(message, claim, "ASSERTS")
memory.add_edge(claim, alice, "ABOUT")

rows = memory.query(
    {
        "match": [
            {"node": "s", "labels": ["Session"], "properties": {"user": "demo"}},
            {"edge": "hm", "type": "HAS_MESSAGE", "direction": "out"},
            {"node": "m", "labels": ["Message"]},
            {"edge": "asserts", "type": "ASSERTS", "direction": "out"},
            {"node": "c", "labels": ["Claim"]},
        ],
        "return": ["m", "c"],
        "limit": 20,
    }
)

for row in rows:
    print(memory.get_node(row["c"]).properties["text"])

Persist the memory graph when sessions should survive process restarts.