How ForsightAI works

Not a chatbot. A multi-agent investigation system that dig across the open web, your knowledge base, and cultural tensions — then distills everything into sourced, actionable briefs.

The gap

Generic AI has a memory problem.

When you ask a chatbot about a trend, it pulls from training weights frozen months ago. It doesn't know what dropped last week. It doesn't know which sources you trust. It doesn't remember what you asked yesterday. Every conversation starts from zero.

ForsightAI was built to close that gap. Fresh web search on every turn. A workspace memory that compounds. A Curator that learns your domain. And structured deliverables that survive beyond the chat.

Architecture

One orchestrator. Six specialists.

An orchestrator routes your question to the right agents. They dig in parallel — across the web, your knowledge base, and cultural archives — then synthesize a single, sourced response. You never manage the routing. It happens silently.

Gap Analysis

Diagnoses market forces, competitive gaps, and white spaces before any concept work begins. This is the mandatory first step for every structured deliverable.

Trends Watch

Deep web search across premium sources — dated signals, drops, market numbers, cultural conversations. Never asserts a fact without a source and a date.

Competition

Crosses your internal KB with live competitive positioning, recent drops, and social signals. Deep search available for cross-source synthesis.

Paradoxes

Detects structural cultural tensions — clean × indulgent, science × ritual, longevity × immediacy. Opportunities live at the intersection of contradictions.

Methodology

Validates every deliverable against a strict concept grid before it reaches you. Sourced, dated, counter-intuitive, concrete, actionable. If any cell is weak, it redoes it.

Curator

A meta-agent that runs after every turn. It extracts rules from your corrections, tracks your trusted sources and reference brands, and grows the workspace doctrine.

Workspace memory

It gets smarter. Not just older.

Every conversation enriches a workspace-specific memory layer. The agent doesn't start from scratch — it starts from everything you've already taught it. Four categories, each lockable.

Trusted sources

Outlets, writers, and studies you've cited or that informed earlier answers. The agent prioritizes these in future searches.

Reference brands

Brands the agent anchors its analyses around. Updated automatically as you discuss competitors, or added manually.

Focus categories

Recurring territories sorted by weight. The weight reflects how often a territory came up across your conversations. Set a manual weight to signal importance.

Learned rules

Explicit behaviours captured from your corrections — 'no jargon', 'always source drops', 'respond in French'. The agent reads these each turn.

The Curator

Your workspace's dedicated teacher.

The Curator is a separate agent accessible from your account. It does three things: ingest reference content into your knowledge base, curate the doctrine (rules and methodologies that shape how the main agent thinks), and manage your workspace memory in plain language.

Talk to the Curator in plain language

Add Glossy as a trusted source

✓ Added Glossy to trusted sources.

Make skincare a high-priority category

✓ Added 'skincare' (weight 80). It now dominates focus categories.

Lock my trusted sources — the agent shouldn't touch them

✓ Locked trusted sources. Only you can add or remove items now.

Remember to always cite the source date for trend claims

✓ Added rule: 'always cite the source date for trend claims'. Read on every turn.

Doctrine

The agent reads what you taught it.

The doctrine is a set of Markdown chunks — routing logic, classification rules, output style, sub-agent behaviors, domain methodologies. It's seeded with sensible defaults, then grows as you and the Curator refine it.

The main agent's system prompt only carries the core (role, golden rules, tool list). Everything else lives in the doctrine table and is retrieved on demand via semantic search. This keeps the prompt lightweight and cacheable, while letting the doctrine grow without bloating the context window.

methodology

Domain methodologies — forecast horizons, tension detection, scenario frameworks, naming grammar.

orchestrator_section

Cross-cutting rules for the main agent — classification, routing, output style, golden rules.

sub_agent_core

Behavior of a specific sub-agent — what it does, when to call it, what it returns.

tool_context

Guidance for a specific tool — when to call it, what inputs it expects, edge cases.

Deliverables

From signal to brief. In one conversation.

When you ask for a concept, the agent runs the full pipeline: opportunity diagnostic, three concept angles across cultural tensions, validation against a strict methodology grid, and a structured brief in your side panel. Not a mood board. A defensible, sourced, named concept.

What a concept brief contains

Code name — memorable, 1-3 words
Dated insight — sourced from web search, not training weights
Named cultural tension — two desirable poles that coexist
Hero component — ingredient, format, or technology
Target segment — precise generational or behavioral target
Comparable precedent — "the X of Y" with a specific drop or claim
Defensible white space — where no one currently plays
Target price + channel — grounded in market reality

How a question flows

From your message to the final brief.

You ask

A question, a thesis, or a vague intuition. The orchestrator maps the investigation scope silently.

Agents dig in parallel

Trends watch searches the web. Competition crosses your KB. Paradoxes detect tensions. Gap analysis maps the opportunity.

Orchestrator synthesizes

All agent outputs are merged into a single, sourced response. No English chunks in a French reply. No unsourced numbers.

Deliverable (if requested)

Ready to investigate?

Start a conversation. The agent will dig, source, and deliver.

Start investigating