Watershed Memory

An AI agent helping water teams protect drinking water after wildfire: connecting river readings, unfinished fieldwork, and verified reports to the next action. Three storms. One memory.

Comment

Inspiration

After a wildfire, the next storm can threaten a community’s drinking-water source. Rain can carry ash, sediment, and debris through a watershed—the land that drains into a shared river or reservoir.

The experience of Las Vegas, New Mexico, made this problem concrete: post-fire flooding and debris threatened the city’s water supply. Protecting that source means following a changing situation across repeated storms, measurements, inspections, and team decisions.

Each new reading arrives beside work that is already underway. An inspection may be incomplete. Evidence may be missing. A review may still need attention.

We built Watershed Memory to keep that work connected. Three storms. One memory.

What it does

Watershed Memory gives source-water teams one continuing case connecting environmental observations, agent decisions, and human field work.

It collects official USGS river-flow, rainfall, and turbidity readings, preserves their history, and identifies evidence that deserves another assessment. The agent examines that evidence alongside earlier decisions and field results, then proposes the next work for a person to approve.

Our defining demonstration is simple:

A field inspection is reported as PARTIAL. Its report and supporting evidence are VERIFIED. The inspection is still unfinished.

The agent reads both facts, continues the existing review, and proposes a follow-up inspection. Verification of a report does not silently become completion of the job.

The operator can approve or change a plan, record results, attach evidence, and correct a report while preserving its history. New readings arrive; the team’s work carries forward.

How we built it

We built a Python application around one Strands agent, Amazon Bedrock, and Amazon Bedrock AgentCore Runtime. The accepted field-aware executions used Amazon Nova 2 Lite.

The engineering makes the agent’s decisions dependable:

  • Continuous source collection: scheduled USGS acquisition preserves timestamps, late publications, and corrected readings.
  • Selective assessment: meaningful changes and scheduled follow-ups bring work back for attention; routine updates stay in the record.
  • Structured agent tools: the agent inspects source health, earlier evidence, reviews, field results, and approved locations before proposing work.
  • Persistent case memory: SQLite stores observations, decisions, operator actions, report revisions, and execution receipts outside the agent session.
  • Validated changes: proposals are checked before they commit. Repeated requests recover saved results without duplicating work.
  • Human authority: the agent proposes; people approve, perform, report, and verify field work.

We verified the current workflow both directly through Bedrock and through AgentCore Runtime. The recorded cloud execution completed ten tool steps and produced a follow-up proposal that remained available for human approval.

The public HTTPS workspace gives each visitor an isolated case. Deployment checks confirmed that an approval survives a service restart and that one visitor’s changes do not affect another’s.

Challenges we faced

The hardest problem was preserving meaning across changing evidence and unfinished work.

A corrected observation must retain its history. A corrected field report must not inherit verification intended for its earlier version. A repeated request must not create a second task. And a verified partial inspection must remain partial.

We addressed these challenges with versioned records, explicit workflow states, immutable context for each agent turn, and durable execution receipts. We tested failure and recovery behavior alongside the successful journey, then independently reviewed the implementation and deployment.

What we learned

Useful agent memory must preserve what changed, what people decided, what they actually reported, and what remains unfinished.

We also learned that a successful model response is only one part of a working agent. The decision must reach the right case, retain its evidence, survive interruption, and leave people in control.

That is the achievement we wanted to demonstrate: environmental evidence becoming accountable, continuing work.

Try it

The next storm arrives. The work stays connected.

Built With

Share this project:

Updates

Submission history