[BUG] MCP search_memory_facts fails with 'Unable to serialize unknown type: <class neo4j.time.DateTime>'

[maui314159]

Summary

The MCP server's search_memory_facts tool consistently returns:

Error searching facts: Unable to serialize unknown type: <class 'neo4j.time.DateTime'>

against a Neo4j-backed graph with non-trivial fact density. search_nodes works fine on the same graph; the failure is specific to fact (edge) retrieval.

Reproduction

Tested on:

• graphiti HEAD = 9cdcc93 (Apr 22, 2026)
• Neo4j community 5.x via neo4j:community Apple-container image
• mcp_server/src/utils/formatting.py unchanged from 375023b
• Group has ~5,000 RELATES_TO edges populated by add_episode_bulk runs over the past 2 months

mcp__graphiti-pe__search_memory_facts(
    query="Payday Netchex add-on synergy",
    group_ids=["gcc_deals"],
    max_facts=3,
)
# → {"error": "Error searching facts: Unable to serialize unknown type: <class 'neo4j.time.DateTime'>"}

Failure is consistent across queries; it's not group-specific or query-specific.

Expected

search_memory_facts returns a FactSearchResponse with created_at / valid_at / invalid_at / expired_at rendered as ISO 8601 strings (matching the behavior of search_nodes and get_entity_edge).

Root cause hypothesis

mcp_server/src/utils/formatting.py:format_fact_result calls:

edge.model_dump(mode='json', exclude={'fact_embedding'})

Pydantic's mode='json' correctly serializes datetime.datetime, but not neo4j.time.DateTime. The fact that this fails for fact search but not node search suggests at least one search-path-internal EntityEdge construction is bypassing graphiti_core.edges.get_entity_edge_from_record (which calls parse_db_date to convert neo4j.time.DateTime → datetime).

The proper fix is upstream of the formatter — every EntityEdge constructor in the search path should funnel through get_entity_edge_from_record (or equivalently call parse_db_date on temporal fields). Defensive coercion in the formatter is a reasonable second line of defense at the API boundary.

Why this may have flown under the radar

• format_fact_result was introduced in the Oct 30, 2025 v1.0.0 MCP refactor (feat: MCP Server v1.0.0 - Modular architecture with multi-provider support #1024); the bug surface is ~6 months old.
• search_memory_facts is less commonly invoked than search_nodes in MCP workflows (topical questions vs. relation-shaped questions).
• Smaller / demo graphs may not exercise the buggy edge-construction path.

Proposed fix

PR forthcoming with a defensive fix in format_fact_result that coerces neo4j.time.{DateTime,Date} → native datetime.datetime before model_dump. Description of that PR will note that the right long-term fix is upstream of the formatter.