Skip to Content

Features

Confidence

How much Docklee trusts each memory, and how that trust changes as facts are confirmed or challenged.

Not every memory deserves the same trust. A fact confirmed by two systems is more reliable than a passing remark from one conversation. Docklee gives each memory a confidence that changes over time, and uses it when it ranks recall results.

What raises and lowers confidence

SignalEffect
The source of the memorySome sources are more reliable than others, and the starting confidence reflects that.
How clearly it could be classifiedA memory with an obvious type starts higher than an ambiguous one.
Agreement from another memoryRaises confidence for both memories.
A conflicting memoryLowers confidence, and both memories are flagged for review.
Age without useLowers confidence slowly over time.

Agreement

When two memories state the same fact, even in different words or from different sources, Docklee links them and raises the confidence of both:

MemoryFrom
"Customer prefers dark mode."Your support chat, through Dock
"User setting: dark theme enabled."Your settings service, through the API

Recall then ranks the preference higher than a fact stated only once.

Low confidence goes to review

A memory whose confidence falls too low is flagged for review instead of being dropped. A person decides whether to keep, correct, or remove it. See Review and correction.

Confidence in recall results

The confidence on a recalled record (confidence_score over HTTP) is a score from 0 to 1 for how well that memory answers your query. It combines relevance with the memory's own confidence and the other ranking signals.

records = lynk.recall_sync("Which theme does the customer use?")
trusted = [record for record in records if record.confidence >= 0.6]

Pick a cutoff by testing your own questions. A fixed cutoff that suits one memory space can be too strict for another.

Next steps

You know how trust is scored. See how memories connect to each other next.

On this page