Bug

addToVectorStore iterates over facts extracted by the LLM and passes each one directly to this.embedder.embed(fact) without validating that the value is a string. When the LLM returns a fact as an object (e.g. {"fact": "some text"} instead of "some text"), Ollama's Go server rejects it:

ResponseError: json: cannot unmarshal object into Go struct field EmbeddingRequest.prompt of type string

Root cause

In addToVectorStore (roughly):

for (const fact of facts) {
  const embedding = await this.embedder.embed(fact);  // fact may be an object!
  newMessageEmbeddings[fact] = embedding;
}

facts is parsed from LLM JSON output. The LLM is an untrusted source — it can return any JSON shape. With less instruction-following models (local models like llama3.1:8b, mistral, etc.), the response may be:

[{"fact": "User prefers dark mode"}, {"fact": "User is named Alice"}]

instead of the expected:

["User prefers dark mode", "User is named Alice"]

This non-string object is then passed directly to OllamaEmbedder.embed(), which does:

const response = await this.ollama.embeddings({ model: this.model, prompt: text });

Ollama's Go endpoint type-checks prompt as string, so passing an object causes an unmarshal error.

Impact

• memory.add() throws on every call when using local/small LLMs
• Memory capture (autoCapture in any integration) silently fails or crashes
• No useful error message — the root cause (LLM returned wrong type) is buried
• Only models that reliably follow strict JSON schemas (GPT-4 class) avoid this in practice

Steps to reproduce

1. Use llama3.1:8b (or any local model) as the LLM provider with Ollama
2. Call memory.add([...], { userId: "test" }) with a complex or multi-turn conversation
3. Observe ResponseError: json: cannot unmarshal object into Go struct field EmbeddingRequest.prompt of type string

Expected behavior

addToVectorStore should validate that each fact is a string before passing it to the embedder. If a fact is an object, extract its text value or convert to string rather than crashing.

Suggested fix

for (const rawFact of facts) {
  const fact = typeof rawFact === "string" ? rawFact
    : (rawFact && typeof rawFact === "object" && "fact" in rawFact) ? String(rawFact.fact)
    : JSON.stringify(rawFact ?? "");
  const embedding = await this.embedder.embed(fact);
  newMessageEmbeddings[fact] = embedding;
}

Similarly, OllamaEmbedder.embed() should defensively coerce its input to string.

Environment

• mem0ai version: latest (tested against dist/oss/index.mjs)
• LLM: ollama provider with llama3.1:8b
• Embedder: ollama provider with nomic-embed-text
• Node.js OSS bindings