Secrets AI Clone — Time Machine Built On Versioned Vector Snapshots

Time Machine Is The Feature That Matters

Secrets AI’s signature feature is called Time Machine. You click any past message in a conversation, edit the text, and the AI regenerates every downstream response while the companion still remembers what they learned about you. It sounds simple. It isn’t.

The implementation requires versioned vector-store snapshots per conversation branch. When the user rewinds and edits, you can’t just replay messages — you have to fork the memory graph at that point, regenerate the downstream completions with the new context, and reconcile any “learned facts” that no longer apply because the original message that taught them has been changed.

Almost no clone vendor ships this. The ones that try usually break persistent memory continuity in the process, which is why Secrets reviewers consistently flag Time Machine as the feature that kept them paying through churn cycles. Their entire memory tier ladder (Free → 4x → 6x) hangs off the back of Time Machine working without dropping context.

This clone ships Time Machine working out of the box on the same vector-database backbone Secrets uses (Pinecone, Weaviate, or self-hosted PGVector — your choice). Plus the eight other Secrets signature features. Plus the nine things Secrets explicitly refuses to build.

How Time Machine Actually Works Under The Hood

The mechanic in plain language:

1. Every chat is a directed graph of message nodes. Each node has a parent (the previous message), a content payload, an embedding vector, and a snapshot pointer into the memory store.
2. When the user edits a past message, you don’t mutate the original node — you fork. A new branch is created from the parent of the edited message. The edited content becomes the new node on the new branch.
3. The memory snapshot at that branch point is cloned, not shared. Any facts the companion “learned” downstream of the original message are now isolated to the old branch.
4. The LLM regenerates the next response on the new branch, using the new edited content and the cloned memory snapshot. As the conversation continues, new memory is written to the new branch’s memory snapshot.
5. The user sees a clean conversation that proceeds from their edit. The old branch still exists in storage, so they can also rewind further and edit again — each edit forks again.

Done naively, this scales O(N²) in memory cost per conversation. Done with copy-on-write snapshots (which is what Secrets is using, judging by their response times), it scales roughly O(N log N) and stays cheap.

Tiered Memory Multipliers — Why 4x And 6x Are A Pricing Trick

Secrets’ memory tier ladder is worth studying as a piece of pricing design:

• Free: ~3 conversations of memory. Standard chat memory only.
• Lite ($5.99/mo): Basic long-term memory across recent sessions.
• Plus ($9.99/mo): Topic + tone recall, ~30-day window.
• Premium ($19.99/mo): “4x memory” — full conversation history, persistent across sessions.
• Ultimate ($39.99/mo): “6x memory” — unlimited recall window with “emotional timeline” reasoning.

The pricing trick is calling them “4x” and “6x” instead of describing what they actually unlock. A user who reads “4x” thinks they’re getting four times as much of something. They’re not. They’re getting access to the full conversation history without truncation, which Secrets could easily ship to every tier — but doesn’t, because the multiplier framing converts at a measurably higher rate than “full memory included” would.

This clone ships the exact same architecture (admin-configurable per tier) plus the option to expose memory limits honestly if you want to differentiate on transparency. Most operators copy Secrets’ framing because it works. Some don’t, and they win the “honest pricing” review segment.

Group Chat With Per-Member Private Memory — Why It’s Hard

Secrets’ group chat feature lets a user pull two or three AI companions into the same conversation. Each companion has their own memory of the conversation — including private memories about the user that the other companions don’t share.

Implementation: each companion in the group has their own memory snapshot pointer. When a new message arrives in the group, every companion’s pointer updates with their own slice of the message + context. Companion A “remembers” their flirty exchange with the user even when Companion B is present, but Companion B doesn’t see those memories.

The hard part is preventing context leak. The system prompt for each companion has to carefully scope what they can reference about other companions, and the LLM has to be tuned not to “explain” memories that aren’t theirs. Secrets does this with a context-firewall layer between memory retrieval and completion. We ship the same firewall.

Moments Economics — What Their Token Costs Reveal

Secrets’ Moments token system gives us a fingerprint of their unit economics. The rates are publicly visible:

• Text message: 1–2 Moments. At ~3,000 messages per dollar of LLM cost, this is roughly 33–67% gross margin.
• Image generation: 30–50 Moments. At ~5 cents of GPU time per image (SDXL on a rented H100), this is roughly 75–85% gross margin.
• Voice call: 100 Moments per minute. At ~3 cents per minute of streaming TTS + STT, this is roughly 90% gross margin — their highest-margin product line.
• Video generation: 600 Moments. At ~25 cents of GPU time per clip (WAN / SVD), this is roughly 70–80% gross margin.
• Custom character creation: 1,000–1,500 Moments. Essentially zero marginal cost; this is pure friction-pricing to slow proliferation.

Operators cloning Secrets should pick their default pricing carefully. Match Secrets exactly and you’ll have their margin profile. Undercut on video (the most-criticized “predatory” line — 15,000 Moments at Ultimate buys only 25 videos/month) and you’ll win conversion at slightly lower margin. The admin lets you tune any of these per action.

Their Pure-SPA Architecture Is Leaking SEO

Secrets is built on Next.js with Turbopack, but rendered as a pure client-side SPA. Every route returns the same 221 KB HTML shell. There’s no SSR, no __NEXT_DATA__, no per-route Open Graph variation, and zero JSON-LD structured data anywhere on the site (I checked).

For a B2C consumer product launched in 2024 with marketing budget behind it, this is fine — they’re acquiring through paid social and affiliate, not SEO. For a clone operator who needs organic discovery to compete with Secrets’ brand spend, it’s not fine.

This clone uses Next.js App Router with proper SSR. Per-character pages render fully on the server with character-specific meta tags, Open Graph cards, and Product / Person / FAQPage / Breadcrumb / Organization structured data. Long-tail character searches (e.g., “anime AI girlfriend with mage personality”) hit your pages before Secrets’.

Nine Things We Ship That Secrets Doesn’t

1. Native React Native mobile apps + APK direct download (Secrets is web-only as of 2025; their unconfirmed Q4 2025 mobile launch hasn’t shipped).
2. REST + GraphQL API with white-label theming and Docker / Kubernetes one-command deploy.
3. Anime-first vertical with separate Pony / NoobAI / Animagine pipelines (Secrets has ~2 anime characters in a 94-character library).
4. Male + female + trans + futanari + non-binary companion support (Secrets is straight-male-targeted, female companions only).
5. Persona diversity presets across 6 ethnicities × 5 body types × 4 age bands × 7 archetypes (Secrets is heavily white / slim / 20s by external review consensus).
6. 3-day Ultimate free trial (Secrets has 200 starter Moments — burns out in 5 minutes).
7. 7-day money-back guarantee (Secrets publishes no refund policy).
8. Full JSON-LD structured data on every commerce surface (Secrets has none).
9. Sub-30-second video generation latency (Secrets averages ~90 seconds — direct conversion advantage at the moment users are deciding to spend Moments).

What’s Inside The $9,000

Time Machine with versioned memory snapshots. Tiered memory multipliers configurable per tier. Group chat with per-member private memory and the context-firewall. Group image generation. Five conversation AI models per character (LLM-agnostic — OpenAI, Claude, Gemini, self-hosted Llama / Mistral / Qwen via vLLM). Moments token economy with admin-configurable per-action pricing. Paragraph Mode (long descriptive RP auto-triggers inline image generation). NSFW intensity slider. Multiple branching chats per character. Self-hosted SDXL / Pony / Flux / Juggernaut image pipeline. WAN / SVD video. WebRTC voice + voice cloning (production-ready, not beta). Discreet billing descriptor pattern. 16-language i18n. Hard age verification. Multi-rail adult-safe payments. Native mobile distribution. SSR + structured data. Full source code. Private demo before you pay. 7-day deployment. 6 months free support.

Who Shouldn’t Buy This Clone

If you want to be a customer of Secrets AI rather than operate your own platform, buy a Secrets subscription directly at $5.99–$39.99/mo. This clone is for operators who want to be the platform.

If you want a thousand-character platform with anime IP characters, look at JanitorAI or Chub.ai — different category, different legal posture. Our clone is built around original realistic + anime personas.

If you want a hobby project to launch on weekends, the cheap NSFW chatbot scripts on Codester will technically run on shared hosting. They won’t survive a real payment processor application or a §2257 audit. We sell production stacks for operators planning to be in business in 12 months.