Agent Memory
What 647 teams taught me about agent memory

Most agent memory is a pile of notes. The agent can find something that looks like your question. It cannot say where a fact came from, whether it is still true, or who else knows it. So it gets things wrong with total confidence.
We put that problem in front of builders at Graph Hacks, the online hackathon my team at FalkorDB (graph DB for agents) ran. 7,899 developers signed up. Over four days, 647 teams took the Agent Memory and Coordination track.
I wrote up the strongest projects, and five patterns came out of them. Each one fixes a common mistake in agent memory, and each has a team that showed how.
1. Link every fact to its source
When agents share memory, one bad source becomes many bad actions. An agent reads a poisoned web page and writes a "fact." Other agents trust it and pay refunds or send emails. Deleting the page fixes nothing that already happened.
kintsugi, by Rupam Golui, records every read, write, handoff, and side effect. When a source gets flagged, one traversal finds everything that depended on it, even across handoffs. Then it re-runs only those actions and keeps the rest.
2. Close old facts
Most memory layers only append. Ask for a customer's plan and you get three answers from three different weeks. I made this same point in my Agent Memory Night talk, and this team built it properly.
Relay, by Tharun Kumar, Dharani K, and Thanushrie Satishkumar, treats memory like a ledger. To prove it, they built a fictional internet provider with an AI support desk. The agents never message each other. Each fact in the shared graph records who wrote it, when it was true, and what it replaced.
3. Give each agent its own graph
Multi-agent systems fail at the seams. One agent promises a refund. Another rejects it in a channel the customer never sees. A third picks up the conversation with no idea what the customer believes.
SecondOrder, by Tushar Agarwal, gives each agent a graph of what it believes, plus nested views of what it thinks other agents believe. It ran 10,200 graphs on one instance at about 2.9 ms per event.
4. Store relationships, so the agent can say why
An agent that stores events and skips relationships has to guess when you ask why.
Emberfall, by Hemant Gaikwad, is a tabletop RPG run by three AI agents. In session 3, Grimma the blacksmith refuses to forge a blade for Kael. In session 1, Kael mocked her brother Borin in a tavern she never visited. Click "Why?" and one query lights up that path.
5. Let memory change what the agent does next
This was the most surprising project of the event. crashgraph, by Ujesh Yadav, is a lunar lander with no neural network and no model file. Before each engine decision, it asks memory what happened the last time it was in this situation. After each attempt, it writes down what went wrong.
It learned to land over 3,743 attempts. Wipe the memory and it lands 0 of 30. Restore a GRAPH.COPY snapshot and it lands 26 of 30. That test is the proof. The memory does the work.
What I would do first
All five patterns take a few lines of Cypher. Pick one thing your agent keeps forgetting. Store it as a graph. Then ask where it came from.
The other Graph Hacks track went after a different problem, so I wrote that up too: Same model, double the score.
The full recap on the company blog has the code and demos for each project: 5 agent memory patterns from Graph Hacks.