October's Swarm Community Call brought together several developments that show how Swarm is moving beyond infrastructure into applications people can actually use. Bee 2.8.3 arrived with improvements to storage incentives and network synchronization. Datafund demonstrated how Swarm can provide verifiable records of AI agent memory. And Freedom Browser demonstrated a new approach to livestreaming: broadcasting directly from a mobile phone to Swarm, without a conventional streaming server or gateway in between.
With Devcon 8 just days away, the call also offered a preview of Swarm's presence across the conference, where much of this technology will be demonstrated, tested, and put into the hands of developers and visitors.
Bee 2.8.3: Faster Sampling, Better Synchronization
Released on 15 October, Bee 2.8.3 brought another round of security, protocol, and performance improvements ahead of Devcon.
One of the more substantial improvements concerns the storage incentives sampling process, where several months of optimization have significantly reduced the time required to complete sampling. The team has also introduced protocol changes to help synchronization converge faster when nodes change their storage radius.
Alongside these improvements, developers have been building better testing and observability tools to measure network behavior, particularly synchronization performance and radius changes. These tools should make it easier to identify bottlenecks and evaluate whether protocol changes deliver the intended improvements.
The next development cycle will continue to focus on security audits, synchronization, and simplifying the Bee codebase by removing unnecessary dependencies and moving parts. Several improvements are already awaiting review.
Community Talk I: Datafund Giving AI Memory a Verifiable History
The October call included a special presentation from Gregor Žavcer, director of Datafund, on the ongoing Data Provenance Fellowship and its application to AI agents.
The underlying question is becoming increasingly important: if an AI agent learns from its memory, how can we know whether that memory has been altered?
As agents become more autonomous, the information they retain between tasks increasingly determines what they know and how they behave. That memory can also be manipulated, or rewritten. An agent's account of what it has done is therefore not necessarily a reliable record of what actually happened.
Datafund demonstrated an upcoming version of PLUR.ai, its open-source AI memory project, with the Provenance Toolkit integrated. The system maintains a hash-chained history of agent memory and creates verifiable references on Swarm. If the recorded history is altered without following the expected process, verification can reveal where its integrity was broken.
The demonstration also made an important distinction between proving that a record existed and proving what caused an AI output. A valid provenance record can establish the integrity of a recorded history and show what information was available in an agent's context. It cannot, by itself, establish why a model produced a particular answer.
This is where provenance becomes more useful than a vague promise of trustworthy AI: it provides evidence that can be independently checked, while making clear what that evidence does and does not prove.
Gregor also emphasized that the provenance functionality is optional and designed with local control in mind. The demonstrated version was a preview, with its public release expected in November.
Beyond plur.ai, the Data Provenance Fellowship is intended to make these capabilities reusable across applications, including AI memory systems, datasets, and other forms of machine-readable knowledge. The invitation to developers was straightforward: integrate provenance into applications so their records can be verified rather than merely trusted.