mem0ai/mem0 · error · Error

Invalid memory action

Error message

Invalid memory action: ${memoryAction}

What it means

VertexAIEmbedder.embed() maps an optional memoryAction ('add' | 'update' | 'search') to a Vertex task type via a lookup table. If memoryAction is defined but not a key of embeddingTypes (i.e. anything other than add/update/search), it throws before making any API call. At the type level this is already narrowed, so runtime hits come from untyped callers.

Solutions

1. Pass only 'add', 'update', or 'search', or omit the argument to use the default SEMANTIC_SIMILARITY task
2. Validate/normalize the action string before calling embed() (lowercase it, map synonyms like 'query' -> 'search')
3. Keep the call site in TypeScript so the literal-union type rejects bad values at compile time

Example fix

// before
await embedder.embed(text, action as any); // action from user input

// after
const action = ["add", "update", "search"].includes(raw) ? raw : undefined;
await embedder.embed(text, action);

const MEMORY_ACTIONS = new Set(["add", "update", "search"]);
const action = MEMORY_ACTIONS.has(rawAction) ? (rawAction as "add" | "update" | "search") : undefined;
await embedder.embed(text, action);

function isMemoryAction(v: unknown): v is "add" | "update" | "search" {
  return v === "add" || v === "update" || v === "search";
}
function isInvalidMemoryActionError(err: unknown): boolean {
  return err instanceof Error && err.message.startsWith("Invalid memory action:");
}

try { await embedder.embed(text, action); }
catch (err) {
  if (err instanceof Error && err.message.startsWith("Invalid memory action:")) {
    return embedder.embed(text); // retry with default task type
  }
  throw err;
}

When it happens

Trigger: JavaScript callers (or 'as any' casts) passing values like 'delete', 'ADD', 'query', or '' as the second argument to embed(); memory pipelines forwarding an arbitrary action string from user input or LLM output into the embedder.

Common situations: Calling the lower-level embed() API directly with an action vocabulary that does not match add/update/search; case mismatches; forwarding unvalidated external strings.

Related errors

• Failed to extract embedding values from batch response
• Failed to extract embedding values from response
• Invalid memory action
• dimension mismatch. Expected , got
• values must be finite numbers for Databricks vector search.

AI-assisted analysis of mem0ai/mem0@001c235229 (2026-08-15). Data as JSON: /api/errors/c521e307a7ad41f3. Report an issue: GitHub.

Appendix: source

Thrown at mem0-ts/src/oss/src/embeddings/vertexai.ts:161

    return {
      content: text,
      task_type: taskType,
    };
  }

  async embed(
    text: string,
    memoryAction?: "add" | "update" | "search",
  ): Promise<number[]> {
    await this.initClient();
    if (!this.client || !this.helpers) {
      throw new Error("Client not initialized");
    }

    let embeddingType = "SEMANTIC_SIMILARITY";
    if (memoryAction !== undefined) {
      if (!(memoryAction in this.embeddingTypes)) {
        throw new Error(`Invalid memory action: ${memoryAction}`);
      }
      embeddingType = this.embeddingTypes[memoryAction];
    }

    const instance = this.formatInstance(text, embeddingType);
    const parameters = {
      outputDimensionality: this.embeddingDims,
    };

    const [response] = await this.client.predict({
      endpoint: this.endpoint(),
      instances: [this.helpers.toValue(instance) as any],
      parameters: this.helpers.toValue(parameters) as any,
    });

    if (!response.predictions || response.predictions.length === 0) {
      throw new Error("No predictions returned from Vertex AI");
    }

View on GitHub (pinned to 001c235229)