Skip to main content

Researcher Example

The researcher example (apps/researcher) is a research assistant that works in sprints. It shows long-form memory and session rehydration: each sprint's findings persist as a structured report, and a fresh session pulls back the relevant history before it starts.

How it uses Walrus memory​

Researcher composes each sprint into one structured report and stores it with remember, then generates recall queries from sprint metadata to rebuild context:

Source: apps/researcher 
const fullText =
  `Sprint Report: ${title}\n\n` +
  `${content}\n\n` +
  `References:\n${references}\n\n` +
  `Sources: ${sourceList}`;

const job = await memwal.remember(fullText);
await memwal.waitForRememberJob(job.job_id);
const { results } = await memwal.recall({ query, limit: 5 });

The structured report format matters: because the whole sprint lives in one memory, recall returns complete findings with their references and sources attached, and the assistant can cite where earlier conclusions came from.

Run it locally​

From the repo root:

Source: apps/researcher 
$ pnpm install
$ cp apps/researcher/.env.example apps/researcher/.env

Fill in the required values in apps/researcher/.env: OPENROUTER_API_KEY, POSTGRES_URL (a PostgreSQL database with the pgvector extension), AUTH_SECRET, and the Walrus Memory values from the dashboard (MEMWAL_PRIVATE_KEY, MEMWAL_ACCOUNT_ID, MEMWAL_SERVER_URL). REDIS_URL, BLOB_READ_WRITE_TOKEN, and the Enoki zkLogin variables are optional.

With the environment configured, apply the database migrations and start the app:

Source: apps/researcher 
$ pnpm --filter researcher db:migrate
$ pnpm dev:researcher

The researcher source documents each variable.

How Researcher generates recall queries​

Before a fresh session starts, Researcher asks an LLM to read each sprint's metadata (title, summary, source titles, and tags) and generate 3-5 diverse search queries. The prompt requires the queries to cover different aspects of the sprint, use terms from the summary and tags, and be phrased as semantic search queries rather than questions. The app then runs every query through recallFromMemWal with a limit of 5 results, deduplicates the returned texts, and assembles them into a context block for that sprint. recallFromMemWal reports each result with a relevance score of 1 - r.distance.