Repository navigation
extract_json silently drops memories when an LLM wraps JSON in prose containing braces #5998
Description
Activity
- added a commit that references this issue
on Jul 2, 2026 Hi, I can take this. I’ll add a regression for brace-containing prose around the JSON payload and update the fallback extractor so it returns the valid JSON object instead of the first-brace/last-brace span.
Silent memory loss is especially dangerous because it looks like a model/personality problem later: the agent simply "forgot," but the actual failure happened at extraction time.
For memory systems, I would treat parser failure as a first-class audit event rather than only an exception path. Even if the JSON cannot be recovered, the system should preserve enough metadata to explain what happened:
memory_write_attempted parser: extract_json status: failed | recovered raw_response_hash error_class fallback_usedIn a personal memory vault pattern, I keep raw events separate from derived memory exactly for this reason: if a derived node fails to build, the source event still exists and can be reprocessed later.
Reference shape:
https://github.com/hegu-1/personal-memory-vault-starterFor Mem0, a robust extractor fix is the direct issue, but surfacing "memory extraction failed" to the caller or telemetry would also help. Otherwise local/chatty models can silently create holes in long-term continuity that are very hard to debug afterwards.
What happened
mem0.memory.utils.extract_jsonis the last-resort parser for LLM responses with no code fence. It returns the span from the first{to the last}. When a chatty LLM (common with local Ollama / LM Studio models) wraps its JSON in prose that itself contains braces, that span begins inside the prose and is not valid JSON.json.loadsraises, the outerexceptin_add_to_vector_storeswallows it, and the extracted memories are silently dropped with no error surfaced to the caller.This is the same silent-memory-loss class as #5245 and #5509, and it falls in the scope
tests/test_chatty_llm_parsing.pyalready targets, but the existing tests only cover prose without braces.Steps to reproduce (offline, no API key)
extract_jsonreturns'{about travel}, here is the update: {"memory": [...]}', which fails to parse, so the memory is lost with no error.Expected behavior
The JSON object (
{"memory": [...]}) should be extracted and stored, regardless of brace-containing prose around it.Environment
mem0 main (reproduced on 2.0.10), Python 3.11, any LLM that returns conversational text around the JSON.