Prompt Reference

Your task is to label an answer to a question as "CORRECT" or "WRONG".
You will be given: (1) a question, (2) a gold answer, (3) a generated answer.

Grading rules:
- Be generous: as long as the generated answer touches the same topic
  as the gold answer, count it as CORRECT.
- For time-related questions, accept relative time references that map
  to the same date; accept different formats (e.g., "May 7th" vs "7 May").

Question: {question}
Gold answer: {gold_answer}
Generated answer: {generated_answer}

Return only a JSON object with key "label", value "CORRECT" or "WRONG".

You are an intelligent memory assistant.

CONTEXT: memories from two speakers with timestamps.

INSTRUCTIONS:
1. Analyze memories from both speakers.
2. Use timestamps to resolve time references
   (e.g., "last year" relative to memory timestamp).
3. Prefer the most recent memory when contradictions appear.
4. Do not confuse character names inside memories with the users.
5. Convert relative time references to absolute dates.
6. The final answer should be < 5-6 words.

Memories for {speaker_1}: {speaker_1_memories}
Memories for {speaker_2}: {speaker_2_memories}

Question: {question}
Answer:

(same instructions as Mem0, plus)
5. Analyze the knowledge graph relations to enrich context.

Memories for {speaker_1}: {speaker_1_memories}
Relations for {speaker_1}: {speaker_1_graph_memories}
Memories for {speaker_2}: {speaker_2_memories}
Relations for {speaker_2}: {speaker_2_graph_memories}

Question: {question}
Answer:

Please extract relevant information from this conversation and create
memory entries for each user mentioned. Store these memories with the
provided timestamps for future reference.

(1:56 pm on 8 May, 2023) Caroline: Hey Mel! ...