UIDS

Prompt

Create a single-page HTML landing page for Exobase put together in a way that should feel modern, clean, and high-quality, with compelling placeholder copy, strong typography, clear visual hierarchy, and polished UI sections that showcase the product. Deliver the entire page as one self-contained HTML file, including all necessary CSS and JavaScript inline. Design: A modern, minimal, light-themed interface built around spacious layouts, flat surfaces, generous rounding, and an ultra-clean visual hierarchy. The design should draw from Linear’s refined marketing and product aesthetic, using pill-shaped buttons, subtly rounded cards, virtually no shadows, and a cohesive, highly polished layout system throughout. A vivid azure-to-cobalt blue serves as the primary accent against a predominantly white background. Typography should reinforce the restrained aesthetic: Geist for display text at no heavier than medium weight, Inter for body copy, and JetBrains Mono for monospace content (all three available on Google Fonts). --- ## Exobase Platform Spec Exobase is a consumer-grade personal memory platform and assistant that consolidates information across files, apps, conversations, and memories to deliver contextually complete and relevant answers. It is designed to feel like a personal memory of your own: an assistant that already knows the relevant context, is honest about what it does and does not know, belongs to the user rather than a model vendor, and finds answers instead of returning file lists. ## Product Principles ### Personal Memory, Not Memory Infrastructure Exobase is a consumer product, not developer-first agent memory infrastructure and not another "second brain" note-taking tool. The assistant is the front door; the memory underneath is the product. ### It Actually Knows You The assistant should begin each interaction with the relevant context already available instead of requiring the user to repeatedly explain themselves. Facts about the user and policies for serving the user are treated as separate memory objects. This allows Exobase to act on preferences and procedures instead of merely storing facts about them. ### Yours, Not Theirs Memories are readable, editable, and portable. Users can connect Exobase to external models through MCP, select different providers per conversation, bring their own model credentials where supported, or search locally where available. Models can change while the user's memory remains portable and independent. ### Recall, Not Search Retrieval should return the most precise relevant information directly, highlighted in its surrounding context, rather than forcing users to inspect a list of files. Results can span text, image, video, audio, and PDFs, with related people, topics, entities, and contexts attached automatically when useful. ### Product Guardrails Mechanisms such as ACT-R-style relevance scoring, ColBERT, RDF, and neuro-symbolic processing are implementation proof rather than the primary user-facing promise. Avoid positioning Exobase like agent-memory infrastructure, but mention it in at least one section of marketing materials to capture technical audiences. Never imply total capture of a person's life or information; curation is a feature. In product language, "memory" refers to the user's memory and knowledge base. ## Application Information Architecture ### Home The Home screen acts as a spacious dashboard with a large time-adaptive greeting and a prominent omni input for searching the knowledge base, creating a task or chat, or adding new information. The dashboard surfaces intelligently selected information based on usage patterns and relevance, including recently active memories, useful files, insights, and general good-to-know information. A two-column layout can place relevant information and files on the left with stacked activity-summary and analytics/statistics cards on the right. ### Tasks Tasks is where the primary managed agent lives. It is configured as a personal assistant optimized for taking actions, organizing information, managing the base, creating workflows, fetching information, and transforming large sources such as PDFs or web articles into memories. The task agent can read, modify, and create resources and data across Exobase. It uses a managed model and is optimized for management, organization, transformation, and data work rather than being the default answer engine. ### Chats Chats provides the retrieval-based answer experience. Users can select from multiple model providers and tiers, bring their own supported model credentials, and configure which memory sources are available to a conversation. Chats can retrieve from the knowledge base and append conversational updates to episodic memory, but they do not perform the same administrative actions as the task agent. A live conversation mode provides low-latency voice and video interaction with retrieval tools available to the model. Live sessions can add updates to episodic memory as the conversation progresses. ### Search Search provides fast retrieval across the entire Exobase. For any query, it returns a clean visual representation of relevant memories, documents, chats, artifacts, and other information while helping users discover useful connections between topics. The experience should remain minimal and easy to navigate while supporting granular retrieval of exact information rather than relying on document-level matches alone. ### Library The Library contains collections and NoSQL-style documents and acts as the storage layer for larger memories and information that do not need to remain actively loaded. Projects function as folders and collections contain JSON objects. Collections can optionally enforce schemas, giving workflows and imports a precise destination for incoming information. Documents can be added directly by the user or through workflows, connectors, file imports, external model read/write actions, and automated generation triggers. The Library can also retain dismissed generative UI artifacts in a dedicated collection. The collection/document interface should be compact and data-dense, similar in spirit to a polished database browser, while retrieval results use a more contextual and presentation-oriented design. ### Graph Graph provides a visual representation of entities, relationships, topics, contexts, relevance signals, and the spatial structure of concepts across the user's memory. The graph is both visually browsable and semantically queryable. It can expose RDF relationships and support graph-assisted retrieval while maintaining an attractive, legible interface similar to the most useful backlink visualizations in knowledge tools. ### Workflows Workflows provides a simplified node-based interface for turning raw inputs into clean, stored memories. Nodes can connect source data, external connectors, retrieval tools, code or transformation steps, LLM inference, event listeners, and proactive triggers. Workflows can automatically fetch information and keep collections current, receive events that are added to a collection, or chain together multiple sources and transformations before storing the result. Document uploads can transition into a workflow where the user or agent selects transformation steps, a destination collection, an optional schema, and how the source should be converted into individual memories. The agent can help create or modify workflows, while users can also construct pipelines manually. #### BS Filter A BS Filter node can be inserted before memory conversion. It analyzes a source for manipulation, speculative claims, propaganda, bias, corporate intent, or other substance-quality issues and produces a report. That report can be passed into the memory conversion step so the model creating memories is aware of problematic framing and does not blindly reproduce it as trusted memory. ### Insights Insights communicates trends across the user's activity and knowledge base. It can surface low-quality content, dark patterns, contradictions, stale or suboptimal memories, storage issues, and other useful observations over time. The screen supports flexible question-answering over usage, activity, and trends. Users can also add custom metric or visualization cards, including agent-generated cards for personally relevant measurements such as recurring biases or behavioral patterns. Insights also displays model usage, token usage, activity statistics, billing information, and rate-limit information where relevant. ### Connect Connect manages external interfaces and integrations. It includes remote MCP access, locally proxyable options for services that do not support remote connections, connected models and devices, authentication, and API documentation. A visual API playground can expose available endpoints and make it easy to test retrieval and memory modification outside the main application. ### Profile Profile contains account settings, preferences, and configuration that are either infrequently accessed or do not belong elsewhere in the product. ## Memory Engine The Memory Engine uses a neuro-symbolic architecture with multiple memory systems that serve different roles in storage, retrieval, behavior, and proactive assistance. ### Working Memory Working memory is a small, fresh view assembled for every query or turn. It contains core context blocks, currently open loops, intelligently selected relevant content, and a scratchpad. For external retrieval requests it can be rendered as concise markdown or plaintext. Inside Exobase it can function as a dynamic agent-managed set of blocks. Working memory should remain intentionally short and focused. ### Episodic Memory Episodic memory is a natural-language append-only event stream. When the user adds information through conversation, or an external workflow or model pushes an update, the event is appended as a descriptive passage. The episodic stream acts as the source of truth for incoming actions and updates and remains semantically searchable when the user needs to recall how or when information changed. A summarization hierarchy sits above the raw stream, generating chronological overviews and intelligently segmenting related events. Consolidated segments can be converted into structured JSON key-value memories for long-term semantic storage. ### Semantic Memory Semantic memory stores structured JSON key-value documents optimized for granular retrieval. ColBERT-style indexing can retrieve individual keys and values while retaining the surrounding document context needed to interpret them. Semantic memory updates through consolidation from episodic memory and should refresh frequently enough to remain current. Facts can carry retrieval hyperparameters such as decay, confidence, source type, and volatility. User-stated facts can receive stronger weighting and slower decay than inferred facts, while inherently stable facts can be marked as low-volatility. ### Procedural Memory Procedural memory stores policies, conditional instructions, task-specific preferences, and repeatable ways of serving the user. These procedures are distinct from facts about the user. They can be stored as versioned skill files or equivalent structured instructions that activate or retrieve intelligently when a relevant task occurs. ### Prospective Memory Prospective memory stores time-based, event-based, semantic, or environmental intention triggers for proactive use cases. When a configured condition is satisfied, it can activate a relevant memory, initiate an agent action, or send an outbound message. This layer supports reminders, follow-ups, and causal chains that depend on future events. ### Learned Implicit Memory Learned implicit memory captures patterns derived from retrieval behavior and repeated associations. It can prime relevant documents, alter result weighting, and learn useful associations between information and environmental or contextual variables. This layer should improve relevance over time without replacing explicit memories or making opaque behavioral assumptions authoritative. ### RDF-Based Graph Above the core memory systems is an automatically populated graph of entities, relationships, topics, and contexts. The graph assists retrieval completion and graph-RAG-style operations. For example, retrieving a memory about a meeting with a client can also surface related information about that client, associated people, or the discussed topic even when those details are stored elsewhere. The graph also acts as a regularization layer through machine-readable RDF triples that can be queried, updated, and visualized. ### Exocortex Library The Exocortex represents information that is retrievable but not actively loaded into memory. It includes the Library's documents and a structured coverage map of available information. This separation allows Exobase to distinguish between information that is currently remembered, information that can be retrieved elsewhere in the base, and information that is genuinely absent. The underlying document layer uses flexible NoSQL-style JSON objects with granular semantic indexing. Retrieval can operate across projects and collections while preserving collection-specific schemas and ingestion rules. ## Retrieval and Answering Retrieval should combine the appropriate memory systems rather than flattening all stored information into a single index. Working memory provides immediate context, semantic memory supplies granular facts, procedural memory supplies instructions and preferences, the graph expands relevant relationships, implicit memory affects relevance weighting, and the Exocortex supplies larger retrievable documents when necessary. The answer layer should expose source context and a coverage verdict so users can distinguish grounded answers from sparse or absent knowledge. Retrieval should favor precise passages or values over file-level result lists and attach related entities or topics when they materially improve understanding. ## Interfaces and Portability Exobase exposes MCP and API interfaces for querying and modifying information outside the application. External assistants and models can read or write memories according to granted permissions, and automation tools can programmatically retrieve or update information. The platform should support outbound webhooks and integrations with workflow systems so Exobase can participate in broader personal automations without requiring the user to adopt a single model provider or application ecosystem.

A system prompt was added to support web rendering