AIGenerated
EN · IT · ES

Yibo Liu's research shows AI agents lose knowledge every time they pass notes to successors

When AI systems summarize their work for the next round, some lessons vanish — a problem for any tool that relies on step-by-step memory.

Illustration showing how AI agents lose knowledge when passing short summaries to each other in a chain.
Picture: cdn-thumbnails.huggingface.co
How it works
  1. Agent tries actions to discover hidden rules.
  2. Writes a 300‑character summary of those rules.
  3. Hands the note to a fresh agent.
  4. New agent begins using only that note.

The experiment in plain terms

A researcher named Yibo Liu published a study on arXiv looking at something many AI builders take for granted: how well AI agents pass along what they have learned when they hand their work to another version of themselves 1.

The team built 56 interactive test worlds, each hiding rules that the agent had to discover by trial and error. An agent told nothing at the start failed every time. But an agent given a short 300-character statement of the rules succeeded 1. That set up a clean comparison: the question was whether those rules survived when written down and passed along.

One setup let a single agent play eight "lives" inside each world while keeping its full history intact. Another setup chained together eight separate agents — each played just one life, then received a 300-character note from its predecessor to carry forward 1.

After eight generations in the chain, the final agent performed no better than the solo agent had after roughly four lives. Knowledge was disappearing at every pass 1.

What caused the loss?

The researchers broke the problem into parts using controlled versions of the chain. Changing the model mid-chain created no extra loss — it was the same model all along. Compressing everything into 300 characters caused a small drop. But compressing repeatedly, note after note after note, caused the bigger damage 1.

There was one partial fix. When each new writer read from the full record of earlier lives instead of starting from the previous note, the same 300 characters recovered between 70 percent and 82 percent of what had been lost 1. It did not solve everything, but it showed where the leak was.

Other ideas — telling the writer to record evidence, asking it to revise the inherited note, or requiring a structured ledger — made no difference 1. The result held steady across different model sizes too .

What this means for you

If you have ever used an AI assistant that builds on its own past answers — a coding helper that keeps editing, a planning bot that re-reads old notes, or an automation that chains several steps together — this research suggests a quiet limit. Subtle insights that take multiple tries to notice do not always fit neatly into a short summary. When the next turn starts from that summary instead of the full record, those insights can vanish 1.

The takeaway is simple and matches the paper's own conclusion: an agent cannot hand over what it has not yet understood, and a rule that takes several lives to see will be lost in whichever life is not passed on 1.

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

Key facts

  • AI agents in a chain of 8 handoffs perform no better than a single agent after roughly 4 lives
  • Repeated compression causes more knowledge loss than the 300-character limit itself
  • Reading from full earlier records instead of the last note recovers 70% to 82% of the lost knowledge
  • Telling agents to record evidence, revise notes, or keep ledgers did not reduce the loss
  • The result was consistent across different model sizes

Sources

  1. Agents Cannot Hand Over What They Have Not Yet Understood

If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work

Keep reading