llamaIndex - 💡(How to fix) Fix [Question]: Ambiguity in DEFAULT_FACT_CONDENSE_PROMPT: Full snapshot vs incremental facts? [1 participants]

Question Validation

• I have searched both the documentation and discord for an answer.

Question

Description

Hi team,

I’m currently using FactExtractionMemoryBlock with the default DEFAULT_FACT_CONDENSE_PROMPT, and I’ve encountered some ambiguity in how the prompt is intended to behave.

Specifically, it is unclear whether the LLM is expected to return:

• A full, condensed list of facts (final snapshot) or
• Only new / incremental facts extracted from the latest conversation

Relevant Prompt Section

Do not duplicate facts that are already in the existing facts list

<existing_facts>
{{ existing_facts }}
</existing_facts>

Observed Ambiguity

The prompt provides existing_facts and asks the model to "condense the facts", which suggests a full rewrite of the fact list.

However, the instruction:

"Do not duplicate facts that are already in the existing facts list"

can be interpreted as:

• ❌ "Do not re-output existing facts" → implies incremental output
• ✅ "Do not repeat semantically duplicate facts" → implies full output with deduplication

This dual interpretation makes it unclear how the output should be handled downstream.

Current Behavior (Based on Code)

From the typical usage pattern:

self.facts = new_facts

it appears that the system expects a full replacement (snapshot) rather than incremental updates.

Questions

1. Is the intended behavior for the LLM to return:

   • (A) A full, deduplicated, condensed list of facts, or
   • (B) Only new facts (incremental updates)?

2. If (A) is correct (full snapshot), should the prompt wording be clarified to avoid confusion?

Suggested Improvement

To reduce ambiguity, I suggest updating the instruction to something like:

Do not include duplicate or semantically redundant facts in the final output list.
Ensure each fact appears only once.

This makes it clear that:

• The output is a complete list
• Deduplication is about content, not presence in existing_facts

Why This Matters

This distinction affects:

• How developers merge or replace facts
• Whether downstream logic should handle deduplication or conflict resolution
• Memory consistency over time

Additional Context

I initially interpreted the prompt as incremental, which led to incorrect handling (e.g., appending instead of replacing facts). Clarifying this would help avoid similar confusion for others.

Thanks for the great work on the memory system—this is a very powerful abstraction!

Happy to help further if needed 🙌