Building AI Agents? Watch Memento (Or Ghajini for a Version with Song and Dance) (testchimp.io)

LLMs acting as agents operate in discrete snapshots—see screen, decide, act, then see a new screen—with no continuous stream of internal context the way humans experience. The team found that giving agents a simple micro-memory prompt changed behavior dramatically: instead of only feeding summaries of past steps, they added one line to the prompt — “Write a note to your future self.” The agent then jots short, actionable reminders (what hypothesis it’s testing, why it chose an action, what to check in the next state), creating a lightweight memory-like thread that carries intent and rationale across otherwise stateless turns. This small prompt-engineering tweak is significant because it produces clearer, more navigable multi-step behavior without complex external memory systems. Technically, it leverages the LLM’s generation ability to persist intent as human-readable notes that improve planning, debugging, and interpretability of agent decisions. Practical implications: easier error diagnosis, better hypothesis-driven exploration, and lower engineering cost than building full retrieval-augmented memory stores. Caveats remain—notes are still ephemeral unless saved, rely on the model following its own record, and don’t replace robust long-term memory architectures—yet as a low-cost intervention, “write a note to your future self” is a powerful way to impart continuity to snapshot-based agents.