fix(mem0-ts): extract content from code blocks instead of removing them

[hunterBhough]

Problem

When an LLM wraps its JSON response in markdown code fences:

\`\`\`json
{"facts": ["The sky is blue."]}
\`\`\`

Explanation: I extracted one fact from the input.

The removeCodeBlocks function strips the fenced block entirely (text.replace(/\``[^\`]*```/g, "")), leaving only the explanation text. The subsequent JSON.parsethen fails withSyntaxError: Unexpected end of JSON input` because it receives:

Explanation: I extracted one fact from the input.

…instead of the actual JSON.

This silently breaks memory.add() — facts are never extracted, and the response shows "No new memories extracted" even though the LLM produced valid output.

Who this affects

Anyone using open-source/local models (Qwen, Llama, Mistral, Phi, etc.) which commonly wrap structured output in code fences even when response_format: { type: "json_object" } is set. This is one of the most common behaviors with local inference servers (llama.cpp, vLLM, Ollama, TabbyAPI, MLX).

Fix

When a code fence is present, extract its content rather than deleting it. Falls back to the original strip behavior if no fenced block is found.

Testing

Tested with Qwen2.5-32B via a local inference server. Before fix: all memory.add() calls silently fail ("no new memories extracted"). After fix: facts are correctly extracted and stored.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[kartik-mem0]

Hey, a more comprehensive fix for this was included in #4044 which covers the code block extraction along with the two-stage JSON parsing. Going to close this in favor of that one — thanks for the PR though.