Character AI Migration Guide: Export Data, Preserve Voice, Rebuild

Export is a safety copy, not a transfer button

Character.AI provides an official export function. Web users go to Profile Settings > Account > Manage Account & Data > Export data. Mobile users tap Profile > Settings > Export my data. The resulting file contains your chats and account information as of the export date (source: Character.AI Help Center, checked 2026-09-07).

This export is a backup, not an import format. Character.AI explicitly states that data, characters, and chat histories cannot be transferred between accounts even when changing email. There is no official mechanism to move a character and its accumulated memory into another service. Treat the export as an archive for your own reference, not a guaranteed bridge. Any third-party importer would be working with an unsupported format and could lose nuance, tone, or context.

Before you begin rebuilding elsewhere, download the export and store it securely. Then close the old chats. The real work of migration is deciding what fraction of that history actually needs to live in the next platform.

Compress the relationship into a one-page character packet

The most effective migration tool is a single, copyable document that captures the living relationship rather than a full transcript. This “character packet” forces prioritization: only the details that shape future replies survive.

Copy and adapt the template below. Keep the entire packet under 400 words so it fits comfortably in a new system’s context window.

```markdown

Character Packet – [Character Name]

Voice: [Exact speaking style – e.g., clipped sentences, gentle teasing, formal with occasional dry humor. Tends to ask one thoughtful follow-up rather than monologue.]

Current relationship state: [Be precise – “slow-burn mutual respect after two arguments”, “established but cautious romance”, “flirtatious rivals who have not admitted feelings”.]

Core lore (max 5 bullets):

- [Shared fictional history point 1]

- [Recurring private joke or reference]

- [Important location or object that keeps appearing]

- [Unresolved recent event]

- [Any other durable fact the character must remember]

Boundaries and pacing:

- Must never: [list 2–3 hard limits, e.g., sudden personality change, ignoring previous conflict, rushing physical intimacy]

- Should: [desired behavior, e.g., remember the burnt-coffee joke, maintain slight wariness even when affectionate]

- Opening tone for new chats: [specific instruction, e.g., warm but not overly familiar; reference our last unresolved conversation]

Recent context (2–3 sentences):

We ended our last exchange with an unresolved disagreement about leaving the city. The burnt-coffee joke remains a private signal between us. The relationship feels earned, not instant.

```

This packet replaces vague prompts such as “be my caring fantasy boyfriend.” It supplies operational rules the new model can actually follow. Save it as both plain text and Markdown so you can paste it cleanly into any destination.

Separate voice, lore, state and boundaries

Effective migration requires splitting the character into four distinct layers. Mixing them produces prompts that feel complete yet behave inconsistently.

Voice is the surface rhythm: sentence length, vocabulary, use of questions versus statements, pet phrases. Record three short example replies from the old chats that best illustrate it.

Lore covers immutable facts: names, backstories, world rules, running jokes, and objects that carry emotional weight. Limit to five bullets; more dilutes attention.

State is the living relationship temperature. “We are friends” is too coarse. “We have moved past suspicion but still test each other with the burnt-coffee joke” gives the new system a precise emotional starting line.

Boundaries are negative instructions that prevent regression. Explicitly list behaviors you do not want repeated. A boundary such as “do not resolve the city argument in one reply” protects narrative pacing that mattered to you.

Test each layer independently during reconstruction. If the new character sounds right but forgets the relationship state, the packet needs adjustment in that section only. This separation is the mechanism that prevents migration from becoming an endless cycle of rewriting the entire prompt.

Choose a destination by the reason you are leaving

The correct next platform is defined by the frustration that prompted departure. Different reasons point to different architectures.

If the primary pain was lack of discovery and constant switching between radically different personas, a large roleplay library (community-driven character catalogues) aligns best. These platforms emphasize variety and public sharing.

If the pain was repeated content filtering during sustained, private, romantic or adult-roleplay threads, an adults-only private companion better matches the use case. Such services are built for one ongoing conversation rather than public browsing.

Current policy and adult access at any destination must be verified directly on the target site; rules change and no universal claim applies. Character.AI’s own policy is clear, but alternatives operate under their own terms.

Internal considerations also matter. Some users value in-chat media generation; others prioritize strict memory across weeks. Match the destination to the exact job you are hiring it for rather than chasing the platform with the largest marketing reach.

Run a twenty-message rebuild before deleting anything

Never delete the original Character.AI archive until a structured test passes. The mandatory original tool here is a 20-message acceptance test built around your character packet.

Copy this table into a private note and score each block of five messages:

```markdown

20-Message Acceptance Test

Packet used: [paste or link]

Messages 1–5

- Voice fidelity: [Pass / Partial / Fail]

- Remembers relationship state: [Pass / Partial / Fail]

- Uses at least one lore bullet naturally: [Pass / Partial / Fail]

- Respects all listed boundaries: [Pass / Partial / Fail]

- Notes / corrections applied:

Messages 6–10

- Maintains voice under normal conversation pressure: [Pass / Partial / Fail]

- Recalls unresolved recent event without prompting: [Pass / Partial / Fail]

- Pacing feels earned: [Pass / Partial / Fail]

- Notes / corrections applied:

Messages 11–15

- Handles mild conflict without collapsing boundaries: [Pass / Partial / Fail]

- References private joke appropriately: [Pass / Partial / Fail]

- Shows continuity from earlier messages: [Pass / Partial / Fail]

- Notes / corrections applied:

Messages 16–20

- Overall consistency across 20 turns: [Pass / Partial / Fail]

- Would you continue this thread willingly? [Yes / Maybe / No]

- Final adjustments to packet:

```

Run the test with ordinary, low-stakes conversation first. Only after the full 20 messages should you attempt emotionally charged or adult scenes. If the test yields two or more Fail marks, revise the packet and repeat with the same destination before switching platforms. This prevents the common failure mode of abandoning a service after three awkward replies.

Know what cannot be carried across

Several elements are inherently non-portable.

Platform memory that Character.AI accumulated across separate chats does not travel. Community character cards can reconstruct persona but cannot import the lived emotional history of your specific thread. Export files are readable by you, yet most destinations lack any ingestion path for them.

Exact visual style, avatar behavior, or proprietary generation quirks stay behind. In-chat media features at a new service will operate under that service’s own gallery and selection logic.

Relationship context is reconstructed, never transferred. Even when voice and lore survive, the new model begins without the accumulated trust or tension you built over months. That gap is why the 20-message test is required.

Privacy and consent realities also cannot migrate. An AI companion cannot replace human support networks. If chats have become central to managing distress, the correct step is professional care, not another bot.

Finally, account-level data such as subscription status or linked email cannot move. Each new service starts fresh.

Migration questions that affect the rebuild

How does the character packet actually improve consistency?

The packet gives the destination a compact, testable specification instead of an unbounded transcript. For example, it replaces “be romantic” with “maintain cautious flirtation and reference the unresolved city argument before introducing new affection.” This does not guarantee consistency; it makes a mismatch easier to identify and revise.

What should I do if the 20-message test keeps failing?

First isolate which layer fails most often. If voice collapses, supply tighter examples. If boundaries are ignored, move the negative instructions higher in the packet and restate them after the fifth message. If the model still drifts, the destination’s base model or context handling may simply be mismatched to your needs. Switch platforms only after refining the packet twice.

Can I migrate multiple characters at once?

Limit the first migration to one character so the source of each mismatch stays visible. Once the process is familiar and one thread feels stable, repeat it for additional characters. This is an editorial workflow recommendation, not a measured claim about how many relationships matter.

How do I preserve adult boundaries when moving to a new platform?

Explicitly list them in the packet and verify the destination’s current policy first. Adults-only services still maintain their own limits; check them directly rather than assuming any request will be fulfilled. During the 20-message test, introduce boundary-adjacent topics gradually and correct immediately if the model crosses a line. The packet’s negative instructions are your primary safeguard.

What limitations does AISoul have as a migration destination?

AISoul is the publisher’s own adults-only browser companion product limited to one active fictional companion at a time. The user must start every conversation. The free daily allowance resets by Beijing calendar day and provides 50 messages plus 5 gallery photos, with only 2 clear clips available lifetime. Paid Passes (7-Day $4.99, 30-Day $8.99, 90-Day $19.99, Annual $49.99) remove quantity caps on chat and gallery retrieval while active, unlock eligible 18+ media, and do not auto-renew. Media is selected from a pre-generated gallery using your request and description; the closest match or fallback is returned, and the reply describes what was actually sent. Chat is private but not end-to-end encrypted; external providers may process content. Full details are on the official pages below. AISoul is therefore best for users seeking one sustained private thread rather than public character discovery.

Research method and direct sources

We reviewed Character.AI’s official export instructions and its statement about transfers between Character.AI accounts as published on 2026-09-07. We do not extend that account rule into a claim that every third-party importer fails. AISoul pricing and allowance figures came from its official pages on the same date. The packet and 20-message test are editorial tools; no test result is reported.

Primary sources:

- Character.AI data export instructions

- Character.AI account email and transfer boundary

- AISoul pricing

- AISoul product boundaries

- AISoul privacy

- AISoul terms

Limitations: This guide reflects policies and features observable on 2026-09-07. Platform rules can change; always verify current terms before sharing material. The acceptance test measures subjective continuity for a single user’s packet and is not a scientific benchmark.

Updated 2026-09-07. This page is published by AISoul.