Memory is one of the features that makes an AI companion feel continuous rather than disposable. But continuity does not come from remembering everything. It comes from remembering the right things at the right time. A practical memory system therefore needs a way to estimate salience: how strongly a past event, preference or relationship detail should influence the next conversation.

Why memory salience is a ranking problem

An AI companion may accumulate thousands of candidate memories: a favorite food mentioned once, a recurring work problem, a family relationship, a temporary travel plan, a sensitive emotional disclosure, or a preference repeated across many sessions. If every item is treated equally, retrieval becomes noisy. The model may surface an old detail when it is no longer relevant, repeat facts the user no longer cares about, or miss a small but emotionally important event because it competes with a much larger archive.

Salience scoring treats memory as a ranking problem. The system asks not only “is this stored?” but also “how important is this for the current relationship and current conversation?”

A useful salience model needs several signals

No single signal is sufficient. Recency matters, but recent facts are not always important. Repetition matters, but repeated small talk should not necessarily outrank a one-time life event. Emotional intensity matters, but it should not cause the system to repeatedly bring up sensitive moments.

A practical score can combine at least five dimensions: recency, frequency, explicit importance, relationship relevance and confidence. Each dimension can be normalized separately before the final ranking step. This makes the system easier to inspect than a single opaque “memory importance” number.

Explicit user signals should outrank weak inference

If a user says, “Please remember that I prefer morning meetings,” that instruction should usually carry more weight than a preference inferred from one conversation. Explicit save, edit and forget actions create high-confidence evidence. Repeated behavior can also increase confidence, but the system should distinguish an observed pattern from a user-confirmed fact.

This is especially important for relationship memory. A companion that confidently states an inferred personal preference as fact can feel intrusive. A safer design stores both the content and the provenance: was it directly stated, inferred, summarized, imported from a profile or created by the system?

Emotional salience needs a ceiling

Emotion is useful because people often expect meaningful experiences to be remembered. A difficult day, an important celebration or a breakthrough in a long-running goal can be more relationship-relevant than routine conversation. However, emotional intensity should not produce unlimited salience. Otherwise the system may keep retrieving painful or private events simply because they were strongly expressed.

One design is to let emotional relevance increase a score only within a capped range, while sensitive categories receive additional retrieval restrictions. This preserves continuity without turning the companion into a machine that repeatedly reopens emotionally charged topics.

Decay should depend on memory type

Different memories should fade at different rates. A temporary travel plan might decay quickly after the trip. A stable preference such as language choice may persist much longer. A relationship milestone may remain available but become less likely to surface unless the context matches.

Decay is not the same as deletion. A memory can remain stored while its retrieval priority gradually falls. That creates a middle ground between “remember forever” and “forget completely.” For a broader treatment of conflicting memories, see our framework on memory conflict resolution in AI companions.

Context should re-rank stored memories

Salience is not fixed. The same memory can be highly relevant in one conversation and irrelevant in another. If the user asks about a job interview, professional goals and recent work context should rise. If the conversation is about travel, location and itinerary memories should become more prominent.

This means retrieval should combine long-term importance with short-term contextual similarity. A high-salience memory that does not match the current topic should not automatically dominate the prompt.

How to evaluate a salience system

Teams can build a small benchmark of relationship scenarios and ask human reviewers whether the retrieved memories are helpful, unnecessary or harmful. Useful metrics include precision of retrieved memories, missed-important-memory rate, stale-memory rate, inappropriate-sensitive-memory rate and the number of repeated memories across consecutive sessions.

The best evaluation is longitudinal. A system may look good in one conversation but become repetitive after weeks of use. Testing should therefore include synthetic relationship histories that span many sessions and contain changing preferences, conflicting facts and temporary plans.

The product goal is selective continuity

A strong AI companion does not need perfect recall. It needs selective continuity: enough memory to make the relationship feel coherent, enough forgetting to avoid clutter, and enough user control to correct the system when it gets something wrong.

Salience scoring is the layer that connects storage with relationship behavior. When it is transparent, contextual and sensitive to confidence, memory becomes more than a database feature. It becomes a mechanism for maintaining identity and continuity without making the user feel that every sentence will be remembered forever.