Improving AI responses by bringing user context into every chat

Owned the full journey, research to shipping through continuous iteration and testing.

Users re-explain themselves every session, pasting the same background into different AI tools, leading to inconsistent outputs. Decisions and context lived in past chats, docs, and threads, but were never reused.

The harder truth: most users don't even know this is happening. They get a shallow response and assume that's just what AI is capable of. They've normalised underperformance, never realising that more context would yield dramatically better results.

When we fed context into a prompt, everything improved. Feeding real user context into a prompt during testing didn't just improve the answer, it delighted users. They weren't expecting the AI to know them that well. That reaction was the signal this was worth building.

Before we could design the injection flow, we had a more fundamental problem: How to bring the context from chats from all the tools (ChatGPT, Claude, Perplexity etc.) into Symmetry?

Two approaches. Both revealed the same gap.

Manual connection worked — but it was high-effort. Users had to connect each chat individually, and the knowledge base would take weeks before it felt useful.

Auto-connection worked too — but it only captured what came next.

While researching parallel products — Nessie, Mem0, Mem, Supermemory — I found that ChatGPT, Claude, and others all offer full data exports. Months of context, sitting in downloadable archives nobody thought to use.

Instead of building from scratch, we let users import what they'd already built.

The flow was built around two things: understanding who the user was, get the import data and filtering out the noise before it ever entered the system. Role and workspace tags did that work quietly in the background.

Two things we were deliberate about:

Transparency: an FAQ explained exactly what was being extracted and what was discarded. Users needed to trust us with their data before handing it over.

Honesty about wait time: extraction wasn't instant. Users got an email when complete; the dashboard showed a progress bar in the meantime. No fake loading screens.

I decided to place the Pill inside the input box — users' focus already lives there, and it was the least intrusive form factor. Injected directly into ChatGPT, Claude, Gemini, and Perplexity, sitting right where the prompt gets written.

The inject button lit up when context was available. Clicking it offered two modes, mapped to where users sat on the trust curve.

Auto Inject pulled all relevant context directly into the input box. One click, no decisions. For users who'd seen enough good picks to trust the system.

Review and Inject opened the Side Panel for users who wanted to see and choose before committing. Neither was the "advanced" option.

The export bet paid off immediately. Every user who completed the import got relevant context injected into their first real prompt. No warm-up period, no empty state to push through. The delight moment we'd seen in early research showed up again in the MVP.

Users graduated from Manual to Auto Inject. New users opened the Side Panel, read the blocks, verified the sources. After a few sessions of good picks they stopped checking. The trust curve we'd designed for played out almost exactly as expected.

Source attribution mattered more than we expected. Seeing which chat or document a block came from removed the hesitation at the inject step. Making the system's reasoning visible turned out to matter more than making the system smarter.

A separate surface from the pill. A circular icon that appeared inside any input field in the browser — Gmail, forms, text editors.

The Composer isn't just injection, it's a full writing layer on top of any text area in the browser. A circular icon appears inside Gmail, forms, or any input field. Three things it can do:

Same underlying system as the pill. Different affordance, broader reach.

The web companion. Users came here to see everything Symmetry had captured, search it, and manage it.

Sequential grouping: docs organized chronologically, by source and session.

Inline editing: correct or annotate without leaving the list view.

Ask Symmetry: conversational search, not just keyword lookup.

Progressive content preview: enough to recognize a doc, not enough to overwhelm.

The dashboard also included multiple workspaces, a folder-based knowledge doc system, a timeline view, and a full Ask Symmetry chat interface — a deliberately thorough system that, in hindsight, I'd have kept much thinner for MVP.