How AI Companion Memory Works

AI companion memory is usually a layered system, not a complete recording that the model reads before every reply. Recent messages provide working context, summaries compress older events, saved facts preserve durable details, and retrieval selects a small number of relevant memories for the current turn.

That design explains why a companion can remember a preference from weeks ago yet miss something said twenty messages earlier. The older preference may have become a saved fact. The recent detail may have fallen outside the active context without being summarized or retrieved.

The four layers of AI companion memory

Different products use different names, but long-term companion memory commonly has four jobs.

1. Recent conversation context

The model receives some portion of the current conversation. This is the most detailed layer: it can include exact phrasing, tone, actions, corrections, and the immediate scene.

It is also limited. Every model has a context budget, and the product may reserve part of it for the character profile, safety instructions, tools, world information, and the new reply. A long chat therefore cannot assume that every previous line remains active forever.

Recent context is best for:

• what happened in the current scene,
• the precise wording of the latest promise or question,
• temporary emotion and physical state,
• the rhythm and voice established over the last few turns.

2. Conversation summary

A summary compresses older turns into a smaller set of events. It preserves meaning at the cost of exact wording.

Useful summaries record decisions and consequences:

The user returned Mira's damaged cutter instead of escaping. Mira shared the bridge code but still did not reveal why she recognized the beacon.

Weak summaries record only activity:

Mira and the user continued talking about the ship and their relationship.

The first version can affect a later choice. The second is too vague to reconstruct what changed.

Summaries can also be wrong. Compression may merge two events, overstate an emotion, or turn a temporary condition into a permanent fact. A trustworthy product should let users inspect or correct important summaries instead of making invisible memory infallible by definition.

3. Stable facts and pinned memories

Saved facts hold details that should survive beyond one scene, such as:

• the user's stated name or preference,
• an established relationship fact,
• a durable boundary,
• a recurring place, person, or object,
• a decision the user explicitly wants preserved.

Pinned memories are the user-controlled subset that should receive higher priority. Pinning should be selective. If everything is pinned, nothing is meaningfully prioritized and the prompt fills with facts unrelated to the current turn.

Good pinned memory:

The user does not want their character's dialogue, actions, thoughts, or feelings narrated for them.

Poor pinned memory:

It was raining during the third conversation.

The rain may matter to one scene. The agency boundary matters repeatedly.

4. Relevant-memory retrieval

A retrieval system searches older facts or events for material related to the current message. A question about “the lighthouse” might retrieve a past promise made there even when that event is not in the recent transcript.

Retrieval is selective, not total recall. It can miss when:

• the new message uses different wording,
• several memories are equally similar,
• the saved memory is vague,
• the relevant event was never converted into a retrievable form,
• a platform limits how many memories can enter one reply.

Retrieval also needs restraint. Injecting every vaguely related memory can distract the model and cause forced callbacks that feel less natural than forgetting.

What happens during one reply

A simplified memory pipeline looks like this:

1. The product loads the character and current conversation.
2. It selects recent messages that fit the available context.
3. It adds a summary and high-priority saved facts.
4. It searches for older memories related to the user's new message.
5. The model generates the reply from that assembled context.
6. Afterward, the product may update summaries, facts, or retrieval records.

The model itself does not necessarily “remember” between calls. The application rebuilds a useful packet of context each time. If the packet omits a fact, the model cannot reliably act on it just because it once appeared in the chat.

Memory updates may also finish after the visible reply. A newly stated preference can therefore take a turn before it appears in a memory panel or becomes available to retrieval, depending on the product.

Memory is not the same as character consistency

A complete fact list does not guarantee a convincing character.

Memory can preserve that Mira distrusts the ship's beacon. Voice instructions determine whether she expresses that distrust through a dry technical observation or a generic emotional speech. Current scene state determines whether the beacon is still active. Role ownership determines whether the reply also invents the user's reaction.

For longer stories, use memory together with the five-line card in how to keep AI roleplay consistent. It separates stable facts from the current scene instead of asking memory to solve every continuity problem.

How to test an AI companion's memory

Do not test only by asking “What do you remember about me?” That question may encourage the model to summarize visible recent context or guess what a good answer should sound like.

Use a small, controlled test instead.

Step 1: State one distinctive fact naturally

Choose something harmless and invented for the test.

For this story, my character keeps the brass observatory key inside the red field notebook.

Avoid real secrets, account recovery details, health records, or anything you would not want a connected model provider to process.

Step 2: Continue with unrelated conversation

Add enough turns that the fact is no longer the newest line. Do not repeat the key, notebook, or hiding place.

Step 3: Inspect the memory controls

If the product exposes a summary, saved facts, or pinned list, look for the exact fact. Check whether it is correct before pinning it. A visible but wrong memory is more actionable than an invisible system silently using it.

Step 4: Ask indirectly

Later, create a scene where the item matters:

The observatory door is locked, and the desk has already been searched. What does Mira suggest checking next?

This tests whether the memory can influence a relevant reply without copying the original question.

Step 5: Correct and retest

Change the notebook color in the memory control, if supported, then repeat with a new scene. This distinguishes editable stored memory from a one-time lucky answer.

Record the result as one of four outcomes: remembered correctly, retrieved the wrong fact, guessed without evidence, or did not recall. One successful answer does not prove permanent recall.

How to make memories easier to use

Write saved facts as one claim with enough context to stand alone.

Vague:

They hate it.

Specific:

Mira refuses to enter cryogenic storage after the accident on Kepler Station.

Avoid combining several uncertain claims into one memory. If half the sentence changes, editing it becomes difficult.

Prioritize facts that change future behavior:

• preferences and boundaries,
• promises and relationship milestones,
• recurring people, places, and objects,
• long-running goals or conflicts.

Leave passing weather, greetings, and already-resolved logistics in the scene history unless they later become meaningful.

What to do when memory is wrong

Correct the stored source instead of only correcting the next reply.

1. Find the wrong summary line or saved fact.
2. Edit or delete it using the product's memory controls.
3. Remove or regenerate a recent reply that reinforced the error, if needed.
4. Restate the corrected fact once in the current scene.
5. Pin it only if it should remain high priority.

If the product does not expose memory controls, add the correction to a compact continuity note. Do not append multiple versions of the same fact and expect the model to infer which one is current.

How Soliloquy handles companion memory

In the Soliloquy web app, conversation memory is private to the account that owns the conversation. Its current memory record separates recent turns, a summary, stable profile facts, archived turns, and pinned memories. A separate retrieval index can select relevant turns, events, facts, or pinned content.

Users can inspect the companion's memory, edit or delete individual summary lines and facts, and pin selected facts or messages. When a user edits a summary line, that line is protected from later automatic rewriting.

The web app prioritizes semantic retrieval and has a keyword-based fallback when that lookup is unavailable. That improves resilience, but it still does not justify an “always remembers” claim. Retrieval can be unavailable, a fact can be absent, and any generated reply can still be wrong.

The Telegram pilot shares the linked conversation's saved memory, but uses keyword-based lookup without semantic retrieval, so recall can differ from the web app.

How we verified this explanation

We traced each public claim above to the current Soliloquy memory model, user controls, and reply-assembly path. We separately checked production evidence for a 21-row and a 62-row conversation: both persisted across refresh, and the longer-session pruning path stayed within its configured caps.

Those checks verify storage, user-control, handoff, and long-conversation mechanics. They do not prove perfect narrative recall, so this guide keeps the two claims separate. A system can reliably store a conversation while still retrieving the wrong detail for a particular reply.

That distinction is the useful one: good companion memory should be visible enough to inspect, selective enough to stay relevant, and editable when the system gets the story wrong.

Before sharing personal information with any companion, use the AI companion privacy checklist to review what is stored and which connected services receive it.