{"slug": "i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway", "title": "I gave my agent the right memory and it ignored it anyway", "summary": "A developer testing a support-agent LLM with memory retrieval found that the agent confidently gave incorrect advice despite having the correct user fact in its context. The agent told an enterprise-plan user to upgrade to enterprise, revealing a failure mode where retrieval succeeds but the response ignores the retrieved memory. The developer notes that popular memory frameworks like Mem0, Zep, and Letta do not verify whether the LLM's response actually reflects the retrieved memory.", "body_md": "A few weeks ago I was testing a support-agent setup — nothing fancy, just\n\nan LLM with a memory layer bolted on so it could remember basic facts\n\nabout a user across sessions. Subscription tier, shipping address, that\n\nkind of thing.\n\nI ran a simple scenario: the user is already on the enterprise plan. I\n\nconfirmed the memory retrieval was working — the fact ** subscription_tier:** came back correctly when I queried \"what tier is the user's\n\nThen I asked the agent, in a support-chat style prompt, what plan the user\n\nwas on.\n\nThe response:\n\n\"Sure, upgrading to our enterprise plan would unlock that feature for\n\nyou.\"\n\nThe user is *already on* enterprise. The agent had the correct fact sitting\n\nright there in its context. It just... used it wrong. Not \"forgot it\" —\n\nthat's a different, more talked-about failure mode. This one is worse in a\n\nspecific way: retrieval succeeded, the fact was injected, and the response\n\nwas still confidently incorrect. Nothing failed loudly. Nothing threw an\n\nerror. If I hadn't been staring at the raw context myself, I'd have had no\n\nway to know this happened except a confused (or annoyed) user telling me\n\nabout it after the fact.\n\nI went looking for how the popular memory frameworks handle this — Mem0,\n\nZep, Letta, the usual suspects. They're all solving real problems: storage,\n\nretrieval, contradiction handling as facts change over time. Zep in\n\nparticular does well on temporal accuracy benchmarks.\n\nBut as far as I can tell, none of them check the thing that actually broke\n\nin my test: *did the LLM's response actually reflect the memory that got\nretrieved for it?* Every framework I looked at seems to assume that once a\n\nSo now I'm curious what other people are seeing. If you're running agents\n\nwith any kind of persistent memory in production —\n\nGenuinely asking — I've been digging into this for a bit and I'm not sure\n\nif I'm looking at something under-discussed or just late to a\n\nwell-known problem.", "url": "https://wpnews.pro/news/i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway", "canonical_source": "https://dev.to/thewilliamboyd93oss/i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway-li7", "published_at": "2026-07-17 01:15:39+00:00", "updated_at": "2026-07-17 02:00:18.671561+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-safety", "ai-research", "developer-tools"], "entities": ["Mem0", "Zep", "Letta"], "alternates": {"html": "https://wpnews.pro/news/i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway", "markdown": "https://wpnews.pro/news/i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway.md", "text": "https://wpnews.pro/news/i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway.txt", "jsonld": "https://wpnews.pro/news/i-gave-my-agent-the-right-memory-and-it-ignored-it-anyway.jsonld"}}