Metamemory: Giving AI the Ability to 'Know What It Knows'
Have you ever been asked a question and known instantly that you do not know the answer? You did not have to search every corner of your brain; your cognitive system performed a rapid "metamemory" check and returned a negative signal.
Spector builds Metamemory into its retrieval engine, allowing agents to diagnose their own knowledge gaps.
The Confidence Assessment Math
When a query is received, Spector executes a pre-retrieval density check on the vector index and semantic graph.
The Metamemory Confidence Score \(MC\) is calculated as:
\[MC = w_1 \cdot \bar{S}_{topK} + w_2 \cdot \ln(\text{Count}_{threshold}) + w_3 \cdot C_{hebbian}\]
Where:
- \(\bar{S}_{topK}\) is the average similarity score of the top-5 retrieved matches.
- \(\text{Count}_{threshold}\) is the number of candidates exceeding a similarity of \(0.7\).
- \(C_{hebbian}\) is the clustering coefficient of the retrieved subgraph.
Hallucination Prevention
If \(MC\) falls below a confidence threshold (e.g., \(MC < 0.35\)), the metamemory system triggers a warning. The agent is notified that it has no reliable historical information on this topic:
Low confidence (MC = 0.18) -> Trigger Metamemory Error -> "I don't have enough information to answer that."
This prevents the agent from hallucinating or generating false answers when history is empty.