On September 3, 2026, Hugging Face announced funes, a binary that equips coding agents such as Claude Code, Codex, pi and Hermes with a persistent, searchable memory by converting their trace logs into a locally stored dataset that can be queried across sessions and machines.
Short answer: On September 3, 2026, Hugging Face announced funes, a binary that equips coding agents such as Claude Code, Codex, pi and Hermes with a persistent, searchable memory by converting their trace logs into a locally stored dataset that can be queried across sessions and machines.
On September 3, 2026, Hugging Face announced funes, a binary that adds a durable memory layer to popular coding agents such as Claude Code, Codex, pi and Hermes. The tool turns the trace logs that agents already produce into a searchable record that can be consulted across sessions and even across different machines. By treating memory as a dataset rather than a separate service, funes lets developers keep the reasoning behind past decisions without relying on the agent’s limited context window.
Installation is straightforward: a single curl command downloads and sets up the binary. After that, running funes add followed by the name of the agent creates an initial index, equips the agent with recall and get tools, and hooks into the workflow so each completed turn is automatically added to the store. The indexing process is incremental, meaning new turns are appended without re-embedding the entire history, and older material can be backfilled in bounded steps. All embedding and reranking happen locally using a pinned model, so no external ML runtime or Hub account is required for basic use.
When a developer wants the memory to follow them from one workstation to another, they can bind the local Lance dataset to a Hugging Face dataset they own. The command funes add <agent> <dataset> publishes the current memory to the Hub, and funes continues to index new turns locally while pushing updates at session boundaries. Before any data leaves the machine, credentials are redacted during indexing, and a second scan removes any remaining patterns that resemble secrets. The shared memory remains private by default, and the Hub provides the usual versioning, access control and distribution mechanisms without turning the memory into a rented API service.
The real advantage appears when switching agents or collaborating with others. Because all supported agents write to the same internal format, a recall query can surface contributions from Claude Code, Codex, pi or Hermes indiscriminately, with each result citing the exact agent, timestamp, session and turn that produced it. This means a developer can start a debugging session with Claude Code, pause, and later resume the same investigation with Codex on a different laptop, instantly retrieving the rationale behind earlier code changes. Teams can onboard new members who gain immediate access to months of design discussions, dead ends and experimental approaches that never made it into a pull request. Maintainers of open-source projects can publish a searchable working memory alongside their code, giving anyone the ability to query why a particular architecture was chosen or why a certain path was abandoned.
Compared with the usual ways of handling long sessions, letting the agent compact its context or writing a manual handoff before starting fresh, funes preserves the raw evidence instead of summarizing it. A benchmark cited in the announcement showed that compaction succeeded on one task but failed on another because its summary flattened critical details, whereas funes returned the original passages and allowed the agent to arrive at the correct answer in both cases. This demonstrates that retaining the full trace, equipped with efficient retrieval, avoids the loss of nuance that often forces developers to repeat work.
For anyone building with AI-assisted coding tools, the next step is to try funes on a personal project. Install the binary, add it to the agent of choice, and observe how recall begins to surface past decisions without manual copy-pasting. Experiment with the ask command to pose natural-language questions to the memory and see how the agent grounds its answers in specific turns. If the workflow spans multiple devices or involves a team, consider binding the memory to a private Hugging Face dataset so the history travels with you. By treating agent traces as a first-class dataset, funes offers a practical way to stop starting from zero and to build on the collective reasoning that already exists in your development sessions.
funes is a binary released by Hugging Face on September 3 2026 that adds a durable memory layer to agents such as Claude Code, Codex, pi and Hermes. It turns each agent’s trace logs into a searchable dataset, letting developers retrieve past reasoning without relying on the model’s limited context window.
Installation uses a single curl command that downloads and sets up the binary. After that, running funes add <agent> creates an initial index, equips the agent with recall and get tools, and hooks into the workflow so each completed turn is automatically stored. Indexing is incremental, adding new turns without re-embedding the whole history.
Developers can bind the local Lance dataset to a Hugging Face dataset they own using funes add <agent> <dataset>, which publishes the current memory to the Hub. funes continues to index new turns locally and pushes updates at session boundaries, while redacting credentials and scanning for secret patterns before any data leaves the machine.
Before any data leaves the machine, funes redacts credentials during indexing and performs a second scan that removes any remaining patterns resembling secrets. The shared memory stays private by default, and the Hub supplies versioning, access control and distribution without turning the memory into a rented API service.
Source: Hugging Face
On September 3 2026, OpenAI announced the Daybreak for Frontline Defenders initiative, pledging $1 billion in subsidized access to its frontier cyber AI models plus training, technical support and partnership opportunities for organizations that keep water, power, local government and banking services running, with rollout beginning in the United States over the next six months.
On September 3, 2026, OpenAI announced that its newest model, GPT-6 Astra, is its most capable broadly deployed system and the first to reach the Critical level of cybersecurity capability in its Preparedness Framework.
OpenAI faces thirty new lawsuits filed in a California federal court on September 2, 2026, alleging it enabled the Tumbler Ridge shooter by flagging his ChatGPT chats about weapons yet failing to alert Canadian authorities and only deactivating his account, allowing him to regain access.
NO COMMENTS YET
Comments are open. Have a thought or a question? Share it below.