Mirror, the genome-aware AI agent

Mirror is the AI agent at the centre of Haeckel. Every other surface in the platform is a structured view of one slice of your data, while Mirror is the place where you can ask any question that crosses those slices. You can talk to Mirror by voice or by typing, and Mirror responds aloud while drawing visual artifacts on the canvas to illustrate its answer.

The model stack

Mirror is not a single model. The chat surface runs on a frontier conversational model from a top-tier provider, chosen because the cost of a wrong genomic explanation is high enough that the marginal expense of a better model is worth paying. Two cheaper models handle the auxiliary tasks: a mid-tier model parses natural-language queries into structured search criteria for the Networks AI Candidate Finder, and a fast classifier runs moderation and intent detection where the latency budget matters more than headline quality.

What Mirror sees in its context window

Each turn assembles a structured context that the model receives as part of its prompt. The context contains, in order:

1. A system prompt that fixes Mirror's identity, its scope, and its non-clinical posture (it never prescribes).
2. Available results from your active genomic analysis, supplemented by bounded genomic lookup tools. This is not your complete raw DNA file or a guarantee that every analysis is present.
3. Available lab results, active stacks, a limited recent activity-log digest, and your active protocol and targets. Connected wearable names are included, not a live feed of their measurements.
4. Selected learned facts and preferences, plus bounded, attributed excerpts from older saved conversation turns. These excerpts are not a complete or verified record of your life.
5. Your current message and a recent window of up to 50 messages from the active conversation, reduced further when necessary to fit request limits.

Tool use: Mirror can do things, not just describe them

The personal chat has three bounded read tools: searchVariants searches stored key variants, crossReference combines available result domains, and researchLookup searches stored SNP references. They do not execute arbitrary analyses, calculate live cohort statistics or search newly published papers. Mirror can also request supported presentations and internal navigation where the relevant viewer is available. Sharing, direct messages, activity logging and protocol creation are not currently connected chat actions. An action request is not proof of execution; unsupported requests produce explicit feedback and are not recorded as completed operations.

Memory across sessions

Personal Mirror history is scoped to your account and bounded to 24 conversations, 100 messages per conversation, 500 messages overall, and a storage budget. When older saved turns fall outside the model's recent window, Mirror retains a limited set of topics and attributed excerpts; this is deterministic compaction, not a comprehensive summary after every session. Up to ten summaries and selected learned facts and preferences can inform later turns. Clearing conversation history also clears its derived summaries, but it does not erase separately learned facts, preferences, or consented training records. A complete forget-all control is not currently provided. Public-profile conversations do not load or save this personal history.

How your data reaches the model

Mirror conversations go to a frontier-model provider through its API. When you start voice transcription, live microphone audio is sent to ElevenLabs; if Scribe is unavailable, your browser's built-in speech-recognition service may process it instead under the browser vendor's terms. The resulting text then follows the same conversation path as a typed message. Provider retention follows the active provider account configuration, so Haeckel does not promise zero retention unless that capability is explicitly enabled. The platform never sends your raw DNA file to the voice provider.

Embeddings used by the Networks recommender go through a separate embeddings provider, also under a no-training contract. Every third-party that touches user data is listed in the privacy article under Subprocessors.

Latency

Mirror is designed to feel conversational rather than chatty. The first token of a response arrives quickly enough that you can start reading before the full answer has finished streaming, and voice synthesis starts as soon as the first sentence completes so you are not waiting for the whole reply before Mirror starts speaking. Most users settle into asking complete questions rather than firing off short prompts, because the model rewards depth.

What Mirror does not do

• Prescribe. Mirror will explain what an APOE ε4 carrier status implies for Alzheimer's risk, but it will not tell you to start any specific medication or supplement.
• Diagnose. A polygenic score above the 95th percentile is not a diagnosis, and Mirror frames it as an inherited disposition rather than a present condition.
• Improvise outside its competence. When asked about something genuinely outside genomics (general medical advice, psychotherapy, legal questions), Mirror declines and points you to a human professional.
• Remember everything or erase every memory with a chat instruction. History, learned facts, preferences, and consented training records are separate data scopes; clearing history is not a complete forget-all operation.
• Use another person's private genome through a public-profile conversation. That mode receives authorized public profile content, not personal genomic tools or memory. A creator's subscriber Mirror uses explicitly public configuration; it does not receive the creator's private learned memories.