Perplexity says Brain, the personalization system behind its Computer agent, now automatically creates and updates reusable skills from users’ memory and past sessions. The goal: retire some of the repeat-yourself routine, turning procedures learned from previous work into instructions for future tasks. SourceSource
It’s on by default for eligible US subscribers and can be turned off in settings, according to the company. But the eligibility instructions could use an edit of their own: the blog names all US Max subscribers, while a November 19 company post says all US consumer and enterprise subscribers. Those statements don’t explain the difference. SourceSource
What “self-improving” actually changes
Brain already maintains a knowledge wiki about a user’s work. Skills add procedures: reusable task instructions, sometimes including scripts, that Computer can load when needed. Perplexity says this personalization changes context and instructions, not the underlying model’s weights. Think an evolving procedure manual, rather than a freshly trained model. Source
During daily runs, the background improvement agent reviews full session records, rather than relying on outcomes or short summaries. Creating a personal skill requires two independent occurrences of the same reusable behavior; updating an existing skill can be justified by one grounded occurrence. Perplexity says the agent abstains when the evidence doesn’t support a change. Not every session needs to become a new house rule. Source
The agent submits proposals rather than directly editing skills. Deterministic checks confirm that proposals are well-formed and applicable, while a separate AI judge audits a sample—not every proposal. Perplexity also caps automatically created personal skills at 10 per user. Adding another at capacity requires retiring an existing skill and justifying why the replacement is more useful. Even the instruction library has a clutter limit. Source
Built-in skills shared across users follow a separate process, using sanitized evidence from sessions eligible for training. A revision must pass every scenario in its evaluation, with the original failing at least one; control scenarios are replayed too. An engineer must approve the change before it goes live. Source
Fewer failures, according to Perplexity
In an eight-week randomized test on production traffic, Perplexity says 72.9% of users with automatic personal-skill improvement enabled received at least one new skill, compared with 6.5% of controls adding one manually. Source
The company reports an 11% relative reduction in the tool-call failure rate across all Computer activity. Excluding the system’s own background runs, the reduction was about 7%. Medium- and high-dissatisfaction threads per user fell 7.6%, and Perplexity says it saw no measurable change in cost or response time. Source
Perplexity doesn’t provide the experiment’s sample size, baseline failure rates or uncertainty estimates. That leaves the absolute failure rate and the precision of the reported gains unclear. Source
For eligible users, Perplexity says the Plugins tab shows available skills, identifies those authored by Brain and lets users choose which are enabled. Source
Comments