Citi launches AI wealth advisor, memory limits remain

Business Insider reports that Citi introduced "Citi Sky," a 24-hour AI-powered wealth advisor, and that the bank plans a phased rollout this summer to certain clients. According to a Citi press release cited by Business Insider, Head of Wealth Andy Sieg said Citi Sky will "change the model of wealth management" and that the tool will become more intuitive "over time." Business Insider reports that Dipendra Malhotra, Citi's head of wealth technology, who participated in developing the tool, said on a panel at New York Fintech Week that a primary limitation for AI agents in wealth management is memory, spanning both short-term and long-term contexts. Malhotra was quoted asking, "One is short-term memory: how long can you have this conversation before you start hallucinations?"

Agentic AI systems used for client-facing finance tasks commonly face memory tradeoffs that affect conversational consistency and factual grounding. Industry-pattern observations note two distinct technical memory challenges: short-term context window management, which governs what the model retains within a live session, and long-term user memory, which covers persistent client facts and preferences across sessions. Companies and research groups address these with hybrid approaches that combine model context, retrieval from external stores, and structured user profiles, but each approach introduces engineering and governance tradeoffs.

Industry reporting and Citi research on agentic AI describe a broader shift toward "do-it-for-me" agents for financial services, where an agent orchestrates data, models, and actions on behalf of clients. For practitioners, that shift raises interoperability, data-provenance, and auditability concerns because financial advice requires traceable data sources and reproducible decision paths. Observed patterns in comparable deployments show that hallucination risk and memory inconsistency are central safety and compliance challenges in regulated domains.

For practitioners: monitor how the rollout manages persistent client state, including access controls, encryption, and consent records. For product teams: watch whether external retrieval mechanisms or structured user databases are used to reduce hallucinations and how those systems are versioned and audited. For regulators and compliance teams: track disclosures, logging, and human-in-the-loop safeguards tied to agent outputs.