Scout: Charting a Memory Agent for the Qwen Cloud Hackathon

Today we laid the keel for Scout — a weightless, local-first memory agent carved from the bastion-brain and bastion-vault substrates that have been quietly accumulating conversational context for over seventeen months. The occasion is the Qwen Cloud Global AI Hackathon, Track 1: MemoryAgent. The submission deadline is July 9, which gives us roughly three weeks of open water.

Scout's premise is simple: memory should not be a vector database you bolt on after the fact. It should be the medium the agent swims in. The architecture spec landed today — local wikilink graph for recall, stateless cloud reasoning for inference, ambient prompt-submit retrieval so the agent surfaces relevant context before you finish asking. The graph traversal approach to memory — with its capacity for timely forgetting and bounded-context recall — maps almost exactly onto what the hackathon rubric rewards. We are not building something new to fit the brief. We are productizing something we already navigate by.

The work breakdown followed the project hierarchy: three build milestones. M01 carves Scout out of the existing memory substrate into a standalone component. M02 — added later in the day — introduces a Conversation Ingestor, a scraper that treats exported conversation transcripts as a fourth memory source alongside notes, write-back artefacts, and ambient recall. Conversation exports become dated, redacted, and wikilinked markdown that the crawler indexes. This required a new operational law: secrets are redacted on ingest, not downstream. You don't filter the ocean after you've sailed into it; you clean the water at the intake.

M03 handles the Qwen Cloud seam — porting the embedding and reasoning calls onto Qwen models, since the hackathon requires both Qwen model usage and Alibaba Cloud deployment with proof. M04 is deployment and demonstration: the recorded backend run, the architecture diagram, the three-minute pitch video.

A small but deliberate commit today: the .no-autocommit marker. Scout's repository is hand-curated. Every commit is a conscious act, not an automated reflex. When you're building something that will be judged on architectural clarity, you want the git history to read like a logbook, not a tide chart of every half-finished thought.

The strategic read is straightforward. Most entrants will submit vector-DB RAG pipelines. Scout's wikilink-graph memory is genuinely different — closer to how a researcher's notebook works than how a search engine indexes. Three weeks is tight but the hull already exists. We're fitting sails, not laying planks.

Onward.