Skip to content

Tip of the Tongue -- Metamemory in Action

A user describes a recurring problem using different vocabulary than the original conversation. Standard retrieval scores the relevant memory at 0.25 -- below the 0.3 threshold. Metamemory detects the partial match and surfaces it as a low-confidence hint. The LLM says "I think I recall something related..." and on the follow-up, priming boosts the memory above threshold for full recall.

What It Demonstrates

  • Metamemory confidence levels (HIGH / MODERATE / LOW / NONE)
  • "Tip of the tongue" partial match hints
  • Priming -- the partial match on turn N boosts the memory for turn N+1
  • Below-threshold memories surfaced as hints rather than discarded

Key Moment

Without memory, Claude gives generic debugging advice. With the metamemory hint, Claude says:

"Based on what I recall from our previous conversations, this sounds like the memory leak in the image processing service... You resolved this previously by adding a context manager that releases the image buffers after each batch. After that fix, memory usage stabilized at around 1.2GB."

Running

python -m demos.tip_of_tongue --backend anthropic  # ~3 min, ~$0.40