The Memory Layer for AI Agents - Drop-in memory infrastructure for AI agents and apps. Context that persists. Built for production.
Python agentic-memoryagentic-memory-systemagentsaiai-agentschatgptgenaillmlong-term-memorymemorymemory-managementpythonragstate-management
fix(oss): validate LLM fact output via FactRetrievalSchema before embedding (#4083)
mgoulartdb15d5c6297e
diff --git a/mem0-ts/src/oss/src/embeddings/ollama.ts b/mem0-ts/src/oss/src/embeddings/ollama.tsindex 41dea7bcf..1bafbeaa9 100644--- a/mem0-ts/src/oss/src/embeddings/ollama.ts+++ b/mem0-ts/src/oss/src/embeddings/ollama.ts@@ -27,9 +27,12 @@ export class OllamaEmbedder implements Embedder {} catch (err) {logger.error(`Error ensuring model exists: ${err}`);}+ // Ollama's Go server requires prompt to be a string. Coerce defensively+ // since callers may pass values parsed from untrusted LLM JSON output.+ const prompt = typeof text === "string" ? text : JSON.stringify(text);const response = await this.ollama.embeddings({model: this.model,- prompt: text,+ prompt,});return response.embedding;}diff --git a/mem0-ts/src/oss/src/memory/index.ts b/mem0-ts/src/oss/src/memory/index.tsindex 5d86e4b3b..46ca8bf2a 100644--- a/mem0-ts/src/oss/src/memory/index.ts+++ b/mem0-ts/src/oss/src/memory/index.ts@@ -15,6 +15,7 @@ import {HistoryManagerFactory,} from "../utils/factory";import {+ FactRetrievalSchema,getFactRetrievalMessages,getUpdateMemoryMessages,parseMessages,@@ -261,7 +262,8 @@ export class Memory {const cleanResponse = removeCodeBlocks(response as string);let facts: string[] = [];try {- facts = JSON.parse(cleanResponse).facts || [];+ const parsed = FactRetrievalSchema.parse(JSON.parse(cleanResponse));+ facts = parsed.facts;} catch (e) {console.error("Failed to parse facts from LLM response:",diff --git a/mem0-ts/src/oss/src/prompts/index.ts b/mem0-ts/src/oss/src/prompts/index.tsindex ef8c79756..e9694b3e5 100644--- a/mem0-ts/src/oss/src/prompts/index.ts+++ b/mem0-ts/src/oss/src/prompts/index.ts@@ -1,9 +1,18 @@import { z } from "zod";+// Accepts a string directly, or an object with a "fact" or "text" key+// (common malformed shapes from smaller LLMs like llama3.1:8b).+const factItem = z.union([+ z.string(),+ z.object({ fact: z.string() }).transform((o) => o.fact),+ z.object({ text: z.string() }).transform((o) => o.text),+]);+// Define Zod schema for fact retrieval outputexport const FactRetrievalSchema = z.object({facts: z- .array(z.string())+ .array(factItem)+ .transform((arr) => arr.filter((s) => s.length > 0)).describe("An array of distinct facts extracted from the conversation."),});