ReportAgent

Inspiration

Teams create weekly reports repeatedly, yet every reporting session starts from zero. Important decisions, preferences, and trends disappear across documents. ReportAgent turns reporting into a learning loop: every report becomes experience that improves the next one.

What it does

ReportAgent accepts dashboard screenshots and notes, uses Qwen Cloud to extract visible metrics, and creates an editable structured report.

Its persistent MemoryAgent:

Embeds every generated report
Retrieves relevant and recent historical experience
Remembers writing style, tone, and focus preferences
Compares trends across reporting periods
Automatically forgets memories outside a configurable retention window
Keeps model context bounded to the five most useful memories

Users can browse report history, ask questions across previous reports, edit generated content, export Markdown or PDF, and listen through Qwen Voice.

How we built it

ReportAgent uses:

Next.js App Router, React, and TypeScript for the full-stack application
Qwen3.7-Plus for text and vision reasoning
text-embedding-v3 and cosine similarity for semantic memory retrieval
qwen3-tts-flash for report narration
Cloudflare Workers KV for persistent production memory
Alibaba Cloud Function Compute deployment specifications
Vitest for automated testing

The memory retrieval system combines semantic relevance, recency, access frequency, and expiration. Recent and semantic results are merged, deduplicated, and limited to five reports before being added to the Qwen context.

Challenges we ran into

The main challenge was making memory useful without flooding the model context. Sending every historical report reduced relevance and increased prompt size.

We solved this by combining recent memories with semantic Top-K retrieval, deduplicating the results, tracking access frequency, and enforcing a strict context limit.

Another challenge was supporting the same persistence interface across local JSON development and serverless KV production.

Accomplishments that we're proud of

Cross-session trend comparisons validated across three weekly reports
Multimodal extraction of dashboard metrics from screenshots
Persistent user style, tone, and focus preferences
Configurable retention and automatic memory forgetting
Memory-grounded chat
Editable reports with Markdown and PDF export
Qwen-generated voice playback
40 automated tests covering APIs, components, memory, and Qwen integration
A live Cloudflare deployment with encrypted API secrets

What we learned

Memory quality depends more on selection than volume. Combining recency with semantic similarity produced more consistent comparisons than sending all historical reports.

Explicit preference memory also makes the agent's learning behavior understandable and controllable by the user.