Useful memory needs time, evidence, and consent
An AI agent becomes more useful when it can distinguish what is current, what changed, what happened repeatedly, and what an owner explicitly prefers. But storing every message as permanent memory would mix casual wording, outdated facts, and private visitor information into the knowledge base.
Qlynk uses a narrower temporal model. Memory is built from approved source claims or explicit owner entries. Each item carries an evidence state, date, confidence, consent basis, and optional expiry so the system can reason about time without treating all history as permanent truth.
Observations, preferences, and patterns are different things
- ObservationA dated event, state, change, behavior, availability note, or preference signal supported by an approved source or owner entry.
- Explicit preferenceA choice the owner deliberately records and can later withdraw, optionally with an expiry date.
- Inferred patternA draft summary supported by multiple dated observations, with a support count and confidence.
- Aggregate intentA privacy-preserving cluster of repeated Knowledge Gaps used for content planning, not personal memory or chat context.
One event does not become a habit
A single observation can be useful history, but it does not prove recurrence. Qlynk requires at least three supporting observations across at least two dates before proposing a recurring pattern. A preference pattern needs at least two dated signals, while an explicit owner preference can be recorded directly because the owner supplied the consent and meaning.
These thresholds do not make an inference certain. They create a minimum evidence boundary. The owner still reviews the proposal, and the agent is instructed to describe a pattern as a tendency rather than a guaranteed future event.
Changes are preserved instead of overwritten
When the same subject and property receive different dated values, the earlier state can matter. A new delivery day, availability window, team owner, or process version should not erase the evidence that explains why an older source or answer was once valid.
Qlynk can propose a change pattern that includes the evidence window and latest value. Retrieval also carries relevant occurrence and expiry dates so the response can preserve phrases such as “as of” or “previously” when the distinction matters.
Expiry and invalidation prevent permanent habits
- Time expiryPatterns receive a bounded lifetime and stop entering retrieval after expiry unless supported and reviewed again.
- Support lossIf current supporting observations fall below the required threshold, an approved pattern becomes unresolved.
- Preference withdrawalAn owner can revoke a preference so it no longer appears in temporal retrieval.
- Source invalidationWhen an approved graph claim loses current source support, the connected observation can become unresolved too.
Private visitor history is not the memory source
Raw conversations, visitor identities, and contact details do not flow into Qlynk temporal memory. The system may group normalized unanswered questions to show that people repeatedly seek a topic, but the result is anonymous aggregate demand and is excluded from the agent’s answer context.
This boundary keeps the focus on improving approved client knowledge. Owners should never copy private visitor records, credentials, payment data, or identity documents into an agent memory entry.