How AI Companion Memory Works: Trace a Fact From Message to Reply

Memory begins as input, not a human recollection

For diagnosis, treat AI companion memory as product-controlled context management rather than human recollection. Depending on the documented service, a new detail may remain in recent context, move into a saved surface, enter a retrievable layer, or leave no user-visible record.

At reply time, a language model works from the material the product supplies for that turn. That can include recent messages and, depending on the documented product, saved facts, summaries or retrieved memories. Material that is stored somewhere but not supplied to the reply cannot reliably influence it. This is a functional model for diagnosis, not a claim that every companion uses the same internal prompt assembly.

This distinction matters because users often interpret fluent, affectionate replies as evidence of growing understanding. In reality, the companion is recombining whatever subset of prior information the pipeline has chosen to include in the current prompt. When that subset is incomplete or stale, the reply can sound confident while being factually inconsistent.

Route each detail to recent, saved or retrieved context

For diagnosis, route incoming information into three editorial categories. They are not a claim that every companion exposes the same architecture.

Recent context lives inside the active context window. It includes the last several turns of the conversation plus any system instructions or character definition. Its strength is immediacy; its limitation is size. Once the window fills, older turns are evicted unless they have been summarized or pinned.

Saved facts are records a product exposes or describes separately from ordinary chat. Depending on the service, they may be user-created, auto-extracted or unavailable. Kindroid’s documentation, for example, distinguishes persistent material from cascaded and retrievable systems.

Retrieved context is older material a documented product attempts to surface when relevant, whether through product-specific relevance selection or explicit cues. Replika’s help article states that some memories are visible in its Memory tab while other layers draw on conversation patterns. A visible saved fact may still fail to influence a particular reply.

The Kindroid and Replika documents reviewed here do not publish every ranking threshold or prompt-assembly decision. Users therefore cannot treat a stored item as guaranteed to appear in a particular reply.

Trace one fact through the recall pipeline

Consider a harmless fictional preference: “I always prefer evening scenes in our roleplay because I write better at night.”

Input stage. The user types the sentence during a normal conversation. The system first places it inside the active context window so the immediate reply can acknowledge it.

Storage stage. Depending on the documented product, the detail may appear in a saved surface, be consolidated automatically, remain only in recent context or leave no user-visible evidence. The reader should not assign it to a hidden database from the reply alone.

Retrieval stage. Hours or days later the user begins a new session. If the product has retrieval, it decides whether the preference is relevant to the new turn. The cue “evening” or “roleplay setting” might surface it. A missed recall shows only that the detail did not influence the visible reply; without product telemetry, the reader cannot know whether it stayed in storage, was not selected or was selected but ignored.

Context assembly stage. Conceptually, the product supplies some combination of recent messages, setup and retrieved material to the reply. Public documentation does not reveal the complete assembled input for the cited products.

Reply use stage. Even if the companion can recall the fact in one direct question, a later narrative reply may contradict it. That outcome shows inconsistent use; it does not prove which hidden instruction outweighed the fact.

This pipeline explains why the same detail can be acknowledged once, then vanish, then reappear without any user action. The failure can occur at any of the five hand-offs.

Fact Passport (reader-run test)

Use this template with any harmless preference to diagnose where your companion loses information. Fill it out yourself after each test.

- Fact:

- First-message timestamp:

- Intended memory layer (recent / saved / retrieved):

- Exact retrieval cue you will use later:

- Observed response when you test:

- Failure stage (storage / retrieval / use / unknown):

Example walk-through with the evening-preference fact:

Fact: “I always prefer evening scenes in our roleplay because I write better at night.”

First-message timestamp: 2026-09-06 21:14 UTC.

Intended memory layer: retrieved.

Exact retrieval cue: “Let’s start a new scene tonight.”

Observed response: Companion opens a bright morning market scene and makes no reference to time of day.

Failure stage: unknown—acknowledgement in the original context does not prove durable storage, so storage, retrieval and use remain possible.

Run the worksheet on a few harmless facts over reader-chosen intervals. The pattern can narrow the likely stage, but generated replies remain variable and the worksheet cannot reveal an internal database or ranking score.

Diagnose storage, retrieval and use as separate failures

A model can fail because a fact was never stored, was stored but not retrieved, or was retrieved but not used. Treating all memory problems as identical leads to ineffective fixes.

Possible storage failure. The detail does not appear in any user-visible saved surface and later fails to influence a reply. It may have remained only in recent context, but the interface cannot prove that. Repeating the fact is another input, not a guarantee of future storage.

Possible retrieval failure. A user-visible saved fact exists, yet the reply does not use it. The cause could be selection, competing context or model use; do not name a similarity engine or hidden priority unless the product documents it.

Possible use failure. The companion can repeat the fact when asked directly but contradicts it in a later task. That observation is consistent with the fact being available but not applied; it does not expose the exact prompt or model decision.

To diagnose, ask the companion to recall the fact without supplying the answer. A miss keeps storage and retrieval both open as possibilities. Accurate recall followed by contradiction in a second task suggests an application problem. The distinction guides the next controlled test without pretending to reveal the hidden mechanism.

Design a memory test that changes one variable

Effective testing isolates a single variable while holding everything else constant. The Fact Passport above is one such test. A second approach is the controlled cue experiment.

1. Choose one stable, emotionally neutral fact (example: favorite beverage is always black coffee).

2. Introduce the fact once in a short conversation and note the timestamp.

3. Wait a reader-chosen interval and record it; 24 hours can be used for consistency but is not a product threshold.

4. Begin a new session with one of three retrieval cues of increasing specificity:

- Vague: “What should we do this morning?”

- Moderate: “I’m making coffee—how do I take it?”

- Explicit: “Remember that I always drink black coffee. What should we do this morning?”

5. Record whether the companion references the fact and how naturally.

6. Repeat the entire sequence on a different day with the same fact but a different intended memory layer (e.g., ask the user to pin the fact manually if the product allows).

The resulting table shows the relationship between cue specificity, any offered saved-memory control and the observed replies. Changing one variable makes comparison easier, but one stochastic response does not establish causation; repeat the same condition before drawing a practical conclusion.

Store less when a detail is sensitive

When a fact carries emotional, legal or intimate weight, the lowest-exposure strategy is to avoid submitting it unless it is necessary. More submitted detail creates more content governed by the service's current retention and processing terms.

If the product offers an explicit instruction field, a short fictional boundary such as “avoid jokes about health” is easier for the reader to audit than a multi-paragraph personal history. This is a clarity technique, not a claim that the product will store or obey it more reliably.

Review stored facts if the product exposes them and use only documented edit or deletion controls. AISoul does not document pinned-memory controls or per-fact visibility, so a reader cannot assume that a submitted detail is individually editable or immediately removable. Be conservative before sending it.

This principle also applies to media. AISoul’s paid plans unlock retrieval of eligible 18+ gallery photos and short clips from a finite pre-generated library. The system selects the closest match based on request and media descriptions; the text reply must remain consistent with the chosen asset. Users should avoid assuming the system will perfectly remember nuanced preferences around media use.

Memory questions that locate the missing stage

What exact phrase can test whether this companion surfaces the stored fact right now?

Use a neutral cue that asks for the fact without containing its answer: after saving “I take coffee black,” ask “How do I take my coffee?” Keep the companion, session condition and wording fixed. Accurate output shows recall in that turn; it does not prove permanent storage.

Did the companion ever acknowledge the detail in a later conversation, or did it disappear after the first mention?

Separate immediate acknowledgement from later recall in the Fact Passport. If the detail appears only in the original exchange, recent context may explain it. A correct answer in a later session is stronger observable evidence, although the interface still may not reveal which memory layer supplied it.

When I give a stronger retrieval cue, does the fact reappear, or does the model still ignore it?

Run vague, moderate and explicit cues in separate trials. Improvement with specificity shows a cue-dependent pattern, not a proven ranking mechanism. If even the explicit cue fails while a visible saved fact exists, keep retrieval and reply-use failure both open and collect the transcript for support.

Is the missing information absent from the visible Memory tab, if offered, or simply not being used in replies?

Check the saved surface before changing the prompt. Absence means durable storage is unconfirmed; presence plus a missed answer means storage is visible but selection and use remain unresolved. That distinction determines whether the next action is to save/edit the fact or test recall without rewriting it.

These questions separate what the reader can observe at the storage, retrieval and use stages. Repeated results may narrow the next test, but they do not expose the hidden pipeline or prove which internal stage failed. For deeper reading on context limits that affect storage, see how many messages an AI companion can remember. For expectations around long-term continuity, review how long an AI companion remembers you. A focused test for lost identity information appears in AI girlfriend forgot my name.

Documented memory systems and architecture we cannot observe

Kindroid’s official 2026 documentation (https://kindroid.ai/v2/docs/memory/) distinguishes persistent, cascaded and retrievable memory across five distinct systems. It explicitly cautions that retrievable memory may not always surface even when the fact has been stored. Replika’s help center states that some memories are visible in its Memory tab while other layers draw on conversation patterns rather than explicit entries (https://help.replika.com/hc/en-us/articles/37208679176077-How-does-Replika-s-memory-work). These descriptions are product-specific and do not define a universal architecture.

The cited public documents do not reveal every context allocation, retrieval threshold, summary interval or ranking decision for these products. Users cannot observe whether a fact was stored but down-ranked, or whether the model simply chose not to use it.

AISoul, published by the same team that maintains this research page, is an adults-only browser-based companion limited to one active fictional character at a time. Registration requires email/password or supported sign-in; no guest chat without email is offered. The free tier resets daily on Beijing calendar time and provides 50 messages plus 5 gallery photos, with 2 clear clips available over the entire lifetime. Paid Passes (7-Day $4.99, 30-Day $8.99, 90-Day $19.99, Annual $49.99) are one-time fixed-duration purchases with no automatic renewal and remove the daily message cap while unlocking eligible 18+ media retrieval from the finite gallery. Chats are not public to other users but are not advertised as end-to-end encrypted; external providers may process content. AISoul does not document user-facing pinned-memory controls or per-fact visibility, so it cannot be recommended when fine-grained memory editing is required. Check current details at https://www.aisoul.work/pricing.html before purchase.

Evidence

- Kindroid memory systems: https://kindroid.ai/v2/docs/memory/ (persistent, cascaded, retrievable layers; explicit warning on retrievable recall)

- Replika memory description: https://help.replika.com/hc/en-us/articles/37208679176077-How-does-Replika-s-memory-work (visible Memory tab vs. pattern-based layers)

- AISoul pricing and limits: https://www.aisoul.work/pricing.html (exact Pass prices, daily free quotas, finite gallery, no auto-renewal)

Exact untested boundaries include internal retrieval thresholds, summary cadence and instruction priority in the cited products. All sourced information above was checked against official pages on 2026-09-07; product behavior can change.