Build agents that remember how work gets done

Watch someone do a real task, document it accurately, and help someone else repeat it. Nine projects for the 42 Paris hackathon, October 16 to 18.

Install Bun 1.3 or newer:

git clone https://github.com/screenpipe/hackathon-starter.git
cd hackathon-starter
bun run demo
bun start

Open http://localhost:4242. No dependencies to install, Screenpipe account or AI key needed for the fictional sample. Review a paid-order reporting procedure, including a refunded order and a one-off browser problem that should not become a standard step.

The bundled candidate is curated fictional data, not proof that AI discovered a workflow. For your own captured work, the optional AI path actually takes observed records as input. It is unapproved output until a person checks it. We have tested the provider adapter with mocked responses; real-provider quality is unverified.

The review fingerprint detects changes to an exported artifact. Reviewer names are self-attested. This is a hackathon prototype, not authenticated enterprise approval, access control or an audit service.

These CLI projects create local Markdown reports. Each includes fictional input, a sample output, acceptance tests and a concrete extension to build. They use deterministic extraction or comparison, so you can inspect the baseline before adding AI.

Try all six with bun run projects. Reports get unique filenames under output/. Use the individual project README for the input contract and commands for your own data. The escalation, handoff and repetition starters require explicit reviewed annotations; they do not infer those fields from raw recordings.

1. In the review desk, compare every instruction with its evidence. Keep the browser refresh classified as troubleshooting, not a normal process step.
2. Answer the open question. For the fictional sample, confirm that pending and refunded rows are excluded.
3. Check each reviewed observation, enter your name, then Approve and freeze. Export the approved JSON to a private location such as local-data/approved.json. Use Export SOP (.md) for a readable handoff with separate normal steps, exceptions and troubleshooting.
4. Read the adapter contract. If it matches your reviewed procedure, run:

bun run replay --approved local-data/approved.json --orders sample-data/orders.csv --adapter paid-order-report-v1
bun run replay --approved local-data/approved.json --orders sample-data/orders-second.csv --adapter paid-order-report-v1

The first result contains Atlas,150.00 and Beacon,80.00. The second contains Atlas,10.30 and Cedar,45.00. Each run writes a new report and receipt under ignored output/, naming the procedure fingerprint, revision and reviewer.

bun run replay --approved local-data/approved.json --orders sample-data/orders-invalid.csv --adapter paid-order-report-v1

The invalid case must reject a duplicate order ID before writing a report. Replay runs fixed code, not AI-generated instructions. Choosing the adapter explicitly is your confirmation that its documented behavior fits the procedure; the prototype cannot verify that semantic match.

Start a new revision in the desk. Its checks reset, and the earlier exported approval remains unchanged. Have a teammate run the instructions without your help and record what they needed clarified.

Follow local-data setup. Capture a small, permissioned work session, inspect it, then:

bun run sop --input local-data/records.json
bun start --draft output/procedure-draft.json

This starts with unclassified observations and unknown context. To ask a configured AI provider for a proposed SOP instead:

bun run sop --input local-data/records.json --ai

Only --ai sends the selected records to that provider, which may charge separately. Unknown evidence IDs and non-matching quotes are rejected, but matching quotes do not establish that an instruction is correct or that the process is complete. The reviewer must make that judgment. No emails, CRM writes or browser actions are executed.

Use fictional or permissioned data for invoice investigation, order exceptions, onboarding handoffs, support escalation or another repeated task. Project ideas and acceptance tests.

The event tracks remain Agents That Remember, Your Data, Your Interface, and Wildcard. These starters give the tracks a concrete starting point. They do not change the organizer's rules or judging criteria.

Optional smaller examples are still available: meeting follow-up and the work-memory browser at /memory.

• Event registration and organizer Discord details. Follow the organizer's #hackathon-access instructions.
• Participant Business access: the 30-day page is being prepared. Organizers will share it after activation is verified. The samples already work without a subscription.
• Install Screenpipe, local API setup, submission checklist.
• Screenpipe MCP, cloud plugin, Gemini CLI extension, community projects.
• Optional advanced River training project, with separate compute/model requirements.

Run bun test and bun run demo to check your fork. Tests run on Mac, Windows and Linux. They cover grounding, review gates, unchanged snapshots, revision lineage, replay on two inputs, project report correctness and failure handling. CI is not evidence of production deployment or independent customer acceptance.

local-data/, output/ and .env are ignored by Git. Exports contain quoted evidence, so inspect them before sharing. Capture and optional cloud services have their own data flows. Original sample code and fictional fixtures use MIT; Screenpipe and linked projects retain their own licenses.