Claude-Mem's latest release adds a privacy check to its server runtime; its tagged code can still leave nested private content and private text in JSON keys, two cases also flagged in a GitHub review. A separate project, Spotter, releases research code for helping robot policies recover from mistakes, with results from simulations and a small physical-robot test.
Claude-Mem v13.36.1, released Oct. 12, says text inside <private> tags could leave the hook in server-runtime events and session-start prompts, then remain in raw server records. The patch strips tagged content before it is sent and checks again before storing events or session metadata. The release says the default worker runtime was not affected.
That protection has limits in the release's tagged code. Its pattern stops at the first closing tag, so a nested private section can leave trailing text behind; its JSON walker checks string values but copies object keys unchanged. A GitHub review bot flagged both cases in the merged PR that added the server check. The repository code at the v13.36.1 tag still has those behaviors. That makes this a partial fix for nested sections and private text in JSON keys, rather than a guarantee that every tagged value is removed from the server path. The repository is Apache-2.0 licensed.
Spotter addresses a different problem: a robot's action policy can make a mistake, such as missing a grasp, and then continue from a bad state. The policy executes short action chunks while a screener checks them in parallel. After warning signs, a vision-language model (a model that interprets images and text) can choose basic arm movements, such as lifting or retreating to an earlier pose, to try to correct the error before handing control back.
In simulation, the paper tested 24 RoboCasa tasks across 1,200 episodes. In the one-example setting, Spotter retrieves a single learned demonstration from separate learning episodes; GPT-6 Astra judges windows flagged by the local screener and checks whether a repair worked. Against each policy alone at 1.8 times its standard task step limit, Spotter raised success by 5.6 percentage points for Cosmos Policy and 7.5 points for π0.5. In a separate RoboTwin 2.0 simulation, π0.5 success rose from 47.2% to 57.0% on the Hard setting. In RoboCasa simulations, successful runs supervised by Qwen took 13 seconds longer with Cosmos Policy and 16 seconds longer with π0.5 than the policy alone at the same 1.8-times step budget. Those successful Qwen-supervised Spotter runs took roughly 70% less time than the Qwen-based Harness VLA runs, which pause policy execution while a vision-language model plans each round. These are simulated results, not physical-robot trials.
The authors also tested a fine-tuned π0.5 policy on a Franka Research 3 robot across three tasks, with 10 trials per task. Success rose from 16 of 30 trials without Spotter to 25 of 30 with Spotter's GPT judge. The test covered one robot and three tasks, so it does not establish performance beyond those settings.
Spotter's repository lists an MIT license and says some adapted files retain Apache-2.0 terms. Both projects are editorial selections, not a ranking or hands-on review; this newsroom did not install or run their code.
Sources
- Claude-Mem v13.36.1 release notes, Oct. 12, 2026
- Claude-Mem repository
- Claude-Mem README at the reviewed commit
- Claude-Mem license at the reviewed commit
- Spotter research preprint, arXiv:2609.36808
- Spotter project repository
- Spotter v1.0 code release, Sept. 29, 2026
- Spotter README at the reviewed commit
- Spotter license at the reviewed commit
- Ryan Kissinger, “Robot Arm,” NIAID BioArt Source / Wikimedia Commons
.png){kind=link}
Image: Ryan Kissinger, courtesy of NIAID; U.S. public domain. The illustration is generic and does not depict Spotter's test hardware. NIAID source.