The previous sessions of different AI coding agents have been combined into one index, and a shared memory "funes" has been released to pass judgments and reasons onto the next task.
Convert an agent-specific history to a common index.
The AI coding agent converts different session records into common indexes that can be searched later on.Hugging Face's work memory software "funes" takes the history of Claude Code, Codex, Pi, and Hermes and searches for past judgments, reasons, and survey results.
The included sessions are then analyzed into a common turn-block format and divided into chunks. The processing system creates each chunk embedded and writes it into the local lance dataset. The agent-specific differences can be absorbed during the analysis phase, so that indexing and search processing can be combined afterwards.
Vector search and select related sessions in BM25
The saved chunk search combines a vector search that deals with the proximity of meaning with the BM25 that deals with the consistency of sentences.
Each search results show which agent the record came from. If a response is needed, funes ask passes the selected session to an external agent such as Claude Code or Codex and generates a response that explicitly indicates the referred session. Funes itself does not have a generated model and is responsible for historical search and transmission.
Multiple agents search for the same work history
The history of Claude Code, Codex, Pi, and Hermes can be indexed into a single memory.By switching between the agents you use, the search scope is not limited to the current product, but you can also make judgments and find out what you’ve done with another agent.
In collaboration with Claude Code, Codex, and Hermes, Hook adds new records to the index every turn. If you are associating shared memory, public processing is also carried out at the session boundary, updating the previous work to be searchable from the next session.
Share local memory as a Hub data set.
Local memory can be published as a dataset in the Hugging Face Hub and can be searched from another device, team member, or other agent.The new dataset repositories created by Funes will be unpublished in their initial state and will continue to be available to existing repositories.
The public case huggingface/funes-memory is not a data set that left the original transcript behind.2million7118It is a derivative index stored on the lance table, and funes recall can specify this remote memory and search it.
Meanwhile, the Hugging Face Dataset Viewer can’t handle the lance format, which is why the huggingface/funes-memory can’t get a partition name, and the browser’s Viewer doesn’t see the dataset.