Features
Ranking
The signals Docklee combines to decide which memories a recall returns, and in what order.
Similar words aren't enough to find the right memory. "What did the customer prefer before the redesign?" needs a memory that's about the customer, from the right time, and still trusted. Docklee ranks every recall on several signals at once.
Five signals
| Signal | Asks |
|---|---|
| Relevance | How closely does this memory answer the question? |
| Validity | Is this memory still current, or has it been replaced? |
| Confidence | How much does Docklee trust this memory? See Confidence. |
| Connections | How many other memories link to this one? |
| Recency | How recently was this memory created or recalled? |
The question shapes the weights
Before ranking, Docklee reads the question and decides which links to follow and which signals matter most.
| The question | Docklee leans on |
|---|---|
| Names a person, company, or product | Every memory about that entity |
| Uses "before", "after", "when", or "timeline" | Memories in time order |
| Uses "why", "because", or "led to" | Cause and effect links |
| Anything else | Meaning |
Meaning is always part of the search, so a recall never returns less than a plain search would.
Ask better questions
| Instead of | Ask |
|---|---|
"maya email" | "How does Maya like to be contacted?" |
"renewal" | "What happened before Northwind renewed?" |
"cancel" | "Why did the customer cancel?" |
records = lynk.recall_sync("What happened before Northwind renewed?", top_k=10)Pass the user's own message as the query. It already carries the names and time words that ranking uses.
Next steps
You know how recall decides. Capture memories automatically next.