Facts not linked to entities, making cross-session retrieval fail

[danstarns]

Problem

When a user saves a fact like "Daniel likes tea" in one chat session, then opens a new chat and asks "what does Daniel like?", the system fails to find any relevant information because no Daniel entity exists to connect the fact to.

Reproduction

Session 1 — User says "Daniel likes tea". The AI calls memory_store (stored as preference):

// Request
{
  "tool": "memory_store",
  "arguments": {
    "memory_type": "preference",
    "content": "Daniel likes tea.",
    "category": "beverage",
    "subject": "Daniel",
    "predicate": "likes",
    "object_value": "tea"
  }
}

// Response — stored successfully, no entity created
{
  "stored": true,
  "type": "preference",
  "id": "f3e0f0c8-e69b-48c7-b8d6-aed29e17656c",
  "category": "beverage"
}

Session 2 (new chat) — User asks "What facts do you know about Daniel?" The AI calls entity_lookup:

// Request
{
  "tool": "entity_lookup",
  "arguments": {
    "name": "Daniel"
  }
}

// Response — entity not found
{
  "found": false,
  "name": "Daniel"
}

The agent responds: "I don't have any specific facts about Daniel. If you want to share something about him, feel free to let me know!" — even though a preference was just stored in the previous session.

Neo4j verification — After Session 1, a MATCH (a) RETURN a query in the Neo4j browser returns only a single node in the entire database:

(:Preference {
  id: "f3e0f0c8-e69b-48c7-b8d6-aed29e17656c",
  preference: "Daniel likes tea.",
  category: "beverage",
  confidence: 1.0,
  created_at: "2026-03-14T03:56:47.344Z",
  embedding: [-0.016, 0.018, ...]  // 1536-dim vector
})

No Entity node for "Daniel" exists. No Fact node exists (the AI chose preference type). The Preference node is completely isolated — no relationships to any other node. This confirms the root cause: memory_store creates the Preference node but never creates or links an Entity node for the subject "Daniel".

Note: The AI chose memory_type: "preference" rather than "fact", but the core bug applies to both — neither add_fact() nor add_preference() create or link Entity nodes. Additionally, the AI in Session 2 only tried entity_lookup and did not attempt memory_search (which might have found the preference via text similarity), highlighting a secondary issue with tool selection.

Expected Behavior

When a fact is stored with a subject (e.g., "Daniel"), the system should:

1. Create or resolve an Entity node for "Daniel" (type: PERSON)
2. Link the Fact to the Entity via a relationship (e.g., ABOUT or RELATES_TO)
3. When a new session queries about "Daniel", the entity lookup finds "Daniel" and traverses to linked facts

Root Cause

1. No entity creation on store: Both add_fact() and add_preference() in LongTermMemory store nodes without extracting or linking entities from the subject/object fields. Facts, preferences, and entities exist as disconnected islands in the graph.

2. entity_lookup has no path to facts/preferences: The MCP entity_lookup tool only searches Entity nodes — it has no traversal path to discover Fact or Preference nodes even when the subject matches.

3. Tool selection gap: The MCP client AI doesn't fall back to memory_search when entity_lookup returns nothing, missing data that could be found via text similarity search.

Proposed Solution

When storing a fact:

• Extract entity references from subject and object fields in both add_fact() and add_preference()
• Resolve against existing entities (or create new ones)
• Create relationships linking the Fact/Preference node to the resolved Entity nodes (e.g., (Entity)-[:SUBJECT_OF]->(Fact), (Fact)-[:OBJECT_OF]->(Entity))
• Ensure entity_lookup and get_context() traverse Fact and Preference relationships when building context
• Consider improving the entity_lookup MCP tool description to hint the AI should fall back to memory_search when no entity is found