Privacy Policy

1. Introduction

Genesys is a causal memory engine for AI agents, operated by Astrix Labs (astrixlabs.ai). This policy explains exactly what data Genesys collects, why, where it is stored, who it is shared with, how long it is kept, and the controls you have over it. It applies to every way you use Genesys: the ChatGPT app, the Claude MCP connector, and the Genesys web app at https://genesys-api.astrixlabs.ai.

We have written this policy to describe what the software actually does, not what it might do in theory. If you find a discrepancy between this document and the product's behavior, please tell us at founder@astrixlabs.ai.

2. What Genesys Is

Genesys stores facts, decisions, preferences, and context you (or an AI agent acting for you) choose to save, and organizes them into a causal graph — a network of memories connected by explicit “this led to that” relationships. Unlike a simple note store, Genesys can explain why a memory matters, trace the chain of reasoning that connects one memory to another, and let you inspect, export, and delete everything it holds about you. Your memories are owned by you, tied to your identity, isolated from other users, and portable across the AI models you connect.

3. Categories of Personal Data We Collect

Genesys is a memory product, so the primary thing it stores is content you or your AI assistant decide to save. We group what we collect into the categories below.

A. Memory content (the core, most sensitive category)

When you use Genesys, you or an AI agent acting on your behalf save “memories.” Each memory can contain free-form text you provide — personal facts, preferences, decisions, project details, lessons learned, corrections, and any other context you choose to store. For each memory we store:

• The full text of the memory and a short summary (the first ~200 characters).
• A topic category label.
• Named entities (people, organizations, places) extracted from the text — only if LLM processing is enabled (see below).
• The source (which assistant saved it — “claude” or “chatgpt”) and the conversation title it came from (note: a conversation title can itself contain personal information).
• Behavioral and scoring metadata used by the memory engine: an importance (“decay”) score, causal weight, how many times the memory has been accessed, timestamps of those accesses, when it was last used, a stability value, and whether you have pinned it.

How memories are created. Memories are stored when you show intent to remember. The assistant saves a memory when you ask it to remember, update, or forget something, when you share information you want kept for later, or when you confirm a durable fact or decision you want retained. The assistant is instructed not to collect personal facts on its own initiative and not to store details you have not indicated you want kept. A few things to understand clearly:

• Genesys only stores what passes through your conversation with the assistant; it does not read your files, email, browsing, or other accounts.
• Everything saved is fully inspectable and deletable by you (see “Your Controls and Rights”). You can see exactly what has been stored and remove any of it.
• If something is saved that you do not want kept, you can delete it at any time or tell the assistant to forget it.

B. Embeddings

To find relevant memories, Genesys converts memory text into numerical vector representations (“embeddings”). These are derived from your content and are stored alongside each memory. Depending on configuration, embeddings are produced either on our servers using a local model or by a third-party API (see below).

C. Causal graph relationships

Genesys connects memories with typed relationship edges (for example, “this decision caused that outcome”). Each edge records the two memories it links, the relationship type and strength, how the link was created (you stated it explicitly, it was auto-linked, or it was inferred by an LLM), and — when LLM processing is enabled — a short natural-language reason explaining the link. Because a reason can mention names or other personal details, our erasure process specifically scrubs personal identifiers from it (see “Your Controls and Rights”).

D. Authentication data

When you connect Genesys, we store a SHA-256 hash of your access token — never the plaintext token — along with your user ID, the OAuth client ID, granted scopes, and an expiry time (24 hours by default). During an active session, plaintext access and refresh tokens exist only in server memory for the life of the process; refresh tokens are never written to disk and are lost on restart.

E. Identity data (via Clerk)

Genesys uses Clerk as its identity provider. We store only your Clerk user ID — the identifier that ties your memories to you. We do not store your password or credentials. When needed (for example, to show your profile in the web app), Genesys fetches your email, name, avatar image, and account-creation date live from Clerk; this profile data is displayed but not saved in the Genesys database.

F. Organization / team data

If you use Genesys within an organization, we store organization records (name, Clerk organization ID, an external billing customer ID, timestamps) and membership records (which users belong to which org and their role), synchronized from Clerk. Memories carry a visibility flag (private or org) so shared memories can be distinguished from personal ones.

G. Rate-limiting and operational counters

To protect the service, Genesys keeps short-lived request counters and live-update subscriber lists in server memory only. These are not written to a database and are cleared on restart. They are used solely to enforce rate limits and stream real-time updates to the web UI.

H. Imported conversation history (only if you use import)

If you choose to import a ChatGPT or Claude conversation export, Genesys parses the archive and turns your messages into memories (storing the full message text) with temporal-sequence links between them. This is a bulk import of your past conversation history and is a significant, sensitive intake. It happens only when you deliberately upload an export file, and imports are subject to a maximum file size.

4. Why We Collect It

We use the data above only for the following purposes:

• To provide the core memory service — saving, retrieving, ranking, and explaining your memories so an AI assistant can recall relevant context.
• To build and traverse the causal graph — linking related memories and letting you and the assistant follow chains of cause and effect (memory_traverse, memory_explain, causal chains).
• To rank relevance — computing embeddings and scores so the most useful memories surface first, and so low-value ones fade.
• To authenticate you and isolate your data — verifying your identity via Clerk and ensuring your memories are visible only to you (or your organization, for org-visible memories).
• To support organizations — managing team membership, roles, and shared (“org-visible”) memories.
• To operate and protect the service — rate limiting, abuse prevention, and reliability.

5. When Your Content Leaves Our Servers (LLM / Embedding Processing)

Two optional processing paths can transmit memory content to third-party AI providers. Whether they are active depends on how the service is configured.

• Embeddings. If configured to use a third-party embedding provider (OpenAI's text-embedding-3-small), the raw text of each memory (truncated to 8,000 characters), including any imported conversation history, is sent to that provider to generate embeddings. Alternatively, Genesys can run a local embedding model on our own servers, in which case your content is not sent to any third party for embedding.
• LLM enrichment. If an Anthropic API key is configured, background workers send memory text (and sometimes pairs of memories) to Anthropic to extract entities, classify categories, generate summaries, infer causal links, and detect contradictions. If this key is not configured, none of these features run and no content is sent to Anthropic.

We disclose these paths because they determine whether your content is processed only by us or also by a third party. Our current production configuration is described in the recipients section below.

6. Where Your Data Is Stored

• Primary store: PostgreSQL with the pgvector extension, hosted on Neon (AWS, us-east-1). This holds your memories, embeddings, causal edges, hashed auth tokens, and organization/membership records.
• Cache: In the production configuration, caching and your core-preference settings are held in-process on the same server — not in an external cache service.
• Hosting / region: The Genesys application runs on Railway infrastructure; the managed Postgres database is hosted by Neon on AWS (us-east-1). The public API is served at https://genesys-api.astrixlabs.ai.

The Genesys codebase includes support for alternative storage backends (MongoDB, a local Obsidian/Markdown vault, and FalkorDB/Redis). These are not used in the production deployment described by this policy. Only the PostgreSQL-on-Neon path above is active for the ChatGPT app, Claude MCP connector, and web app.

7. Recipients and Subprocessors

Your data is disclosed only to the service providers necessary to run Genesys:

• Railway — cloud application hosting. Data in transit through the application (compute layer).
• Neon — managed PostgreSQL database (AWS, us-east-1). Holds all data at rest.
• Cloudflare — CDN, TLS, and security in front of our web properties. Sees request metadata (IP address, headers) in transit.
• Clerk — identity and authentication. Receives authentication interactions; your profile (email, name, avatar) lives at Clerk; organization events are sent to us via webhook.
• OpenAI — embedding generation (only when the third-party embedder is enabled). Receives the raw text of each memory and any imported conversation history, to produce embeddings.
• Anthropic — LLM enrichment (only when enabled by API key). Receives memory text and memory pairs, to extract entities/categories/summaries and infer causal links and contradictions.

• If Genesys is configured to use the local, on-server embedder, OpenAI does not receive your content.
• If no Anthropic key is configured, Anthropic receives nothing.
• An external payment processor (Stripe) identifier may be stored on organization records, but no billing or payment data is processed by Genesys — billing, if any, is handled entirely outside this service. Consistent with platform rules, subscriptions are not sold inside the ChatGPT app.
• We do not share your data with advertisers, data brokers, or any recipient not listed here.

8. Data Retention

We want to be precise here, because retention in Genesys is nuanced and a reviewer or user can test it directly.

The default is: your memories persist until you delete them. Genesys does not have a blanket time-based expiry that erases your memories after N days.

Memory scoring and “forgetting.” Genesys continuously recalculates an importance score for each memory (roughly every 10 minutes). Low scores cause a memory to lose prominence and eventually move to a “dormant” status — but dormant memories are still stored and still retrievable. Scoring is not the same as deletion.

When automatic deletion can happen. A memory is permanently and automatically deleted by the forgetting process only if it meets ALL of these conditions at the same time:

• Its importance score has fallen below the forgetting threshold (very low value), and
• It has no supportive causal connections to other memories (it is an orphan), and
• It is not pinned, and
• It is not a “core” memory, and
• It is not an organization-visible memory.

In other words: a memory that is connected to others, pinned, marked core, or shared with your org is never auto-deleted — it is retained until you delete it. Only isolated, unpinned, low-value memories are pruned automatically. We state this precisely so you can rely on it: we do not claim “we never delete,” and we do not claim “forgetting erases your data” — both would be inaccurate.

Authentication tokens expire within 24 hours and become inert; expired token records are ignored on every lookup.

In-memory operational data (rate-limit counters, live-update subscriptions, session tokens) is transient and cleared when the server restarts.

9. Your Controls and Rights

You can inspect, correct, and delete your data. The controls available depend on the surface you are using.

Available directly to you through the AI assistant (ChatGPT app / Claude MCP)

• Inspect what's stored — search, recall, list your core memories, view statistics, traverse the causal graph, and ask why a memory is retained (memory_search, memory_recall, list_core_memories, memory_stats, memory_traverse, memory_explain). This makes your memory fully transparent and portable.
• Delete a specific memory — delete_memory permanently removes an individual memory; its connecting edges are removed automatically. This is immediate and irreversible.
• Control retention — pin_memory protects a memory from ever being auto-forgotten; unpin_memory removes that protection.
• Control auto-promotion — set_core_preferences lets you choose which categories of memory may be promoted to “core,” which require approval, and which are excluded.

Available through the web app

• View your memories, causal graph, timeline, core memories, and statistics via the web UI.
• Export everything (data portability) — an authenticated request to GET /api/export returns your complete memory graph — every memory (all statuses) and every causal link — as a single JSON document you can save and take with you to any other tool or AI model (GDPR Article 20 portability). Your memory is not locked to Genesys or to any one AI vendor.
• Promote a memory to organization visibility — this is a user-initiated action that shares a selected memory with your organization.

Full account erasure (GDPR Article 17)

Genesys implements a complete erasure operation. Currently, this full erasure is executed by our team on request rather than through a self-service button in the ChatGPT app. To erase your account data, email founder@astrixlabs.ai and we will process it. When we run erasure:

• All of your private memories and their causal edges are permanently deleted.
• Your organization membership records are deleted, and your cached preferences are deleted.
• Memories you previously promoted to your organization are retained but anonymized — your user ID is replaced with an anonymized marker (erased_user) on every retained record, and personal identifiers (emails, names) are scrubbed from the text of connected causal-link explanations. This is by design so shared team knowledge is not destroyed for other members. If you want these fully deleted too, tell us and we will run a complete deletion instead.
• Note that erasure does not need to purge short-lived auth tokens (they self-expire within 24 hours) and does not delete the organization record itself.

We are working to expose self-service full erasure directly in the app. In the meantime, per-memory deletion (through the assistant) and full export (GET /api/export) are available to you at any time, and full erasure is one email away.

You also have the right to access, correct, or object to processing of your data, and to lodge a complaint with your local data protection authority. If you are a California resident, you have the right to know what data we collect, request deletion, and opt out of sale — and we do not sell your personal data. Contact us at founder@astrixlabs.ai to exercise any of these rights.

10. Security Measures

• Per-user isolation. Every memory, edge, and token is tied to your user ID; queries are scoped so you can only access your own data (and organization-visible data for orgs you belong to).
• Hashed tokens. Access tokens are stored only as SHA-256 hashes, never in plaintext. Refresh tokens are never written to persistent storage.
• OAuth 2.0 with PKCE. Authentication uses a Clerk-backed OAuth provider with dynamic client registration and PKCE (S256), and short-lived (24-hour) tokens.
• Transport security. All API traffic is served over HTTPS.
• Managed infrastructure. Data at rest is held on managed PostgreSQL provided by Neon; application compute runs on Railway.

No system is perfectly secure, but we design Genesys so that the most sensitive material — your credentials — is never stored in a form we could read.

11. Children's Privacy

Genesys is not directed to children under 13 (or the minimum age in your jurisdiction), and we do not knowingly collect data from them. If you believe a child has used Genesys, contact us and we will delete the data.

12. International Users

Genesys is operated from, and stores data on, infrastructure that may be located in the United States. If you use Genesys from outside the United States, you understand your data will be processed there and by the subprocessors listed in the recipients section above.

13. Changes to This Policy

We may update this policy as Genesys evolves. When we make a material change, we will update the “Effective date” at the top and, where appropriate, notify you. Continued use of Genesys after an update means you accept the revised policy.

14. Contact

• Astrix Labs
• Email: founder@astrixlabs.ai
• Website: astrixlabs.ai

For any privacy question, data access request, or full account erasure, email us and we will respond promptly.