Agent 'gets smarter' demo

A new April 12 YouTube demo pushed Hermes Agent into the memory debate by showing an assistant that claims it “gets smarter” as people keep using it. (youtube.com) Hermes is an open-source agent from Nous Research, and its own documentation says it uses a “closed learning loop” with persistent memory, autonomous skill creation, skill self-improvement, and cross-session recall. (hermes-agent.nousresearch.com) In plain terms, the pitch is that the agent does not start from zero each time a chat ends. Hermes stores bounded notes in `MEMORY.md` and `USER.md`, keeps them under fixed character limits, and injects them into the next session as context. (hermes-agent.nousresearch.com) Hermes also says it can turn repeated work into reusable “skills,” which are like saved procedures the agent can call again instead of improvising from scratch. Its docs describe that as procedural memory, separate from the user-profile memory that tracks preferences and habits. (hermes-agent.nousresearch.com) The “gets smarter” claim does not mean the underlying language model rewrites its own weights after every chat. Hermes’ public docs frame the improvement as an engineering loop around the model: memory files, search over past sessions, and skills that can be created or refined during use. (hermes-agent.nousresearch.com, github.com) That distinction has become a live issue in 2026 because many agents can use tools, but fewer promise cross-session adaptation that is visible to the user. Hermes now lists 47 built-in tools, 14-plus messaging platforms, and multiple external memory providers that can extend personalization beyond the built-in files. (hermes-agent.nousresearch.com, hermes-agent.nousresearch.com) One of those add-ons, Honcho, is described by Nous as a memory backend that builds a running model of the user’s goals, communication style, and patterns after conversations happen. That is closer to a long-term dossier than a simple chat history. (hermes-agent.nousresearch.com) The project is also moving fast enough that the demo landed days after a major release. GitHub shows Hermes Agent v0.8.0 was published on April 8, 2026, with 209 merged pull requests, 82 resolved issues, and features including background task auto-notifications and live model switching. (github.com, github.com) Interest has followed that pace. GitHub’s releases page showed more than 73,000 stars on April 12, while the main repository snapshot available through search showed 51,300 stars earlier in the project’s recent climb, underscoring how quickly attention has moved. (github.com, github.com) The open question is whether users treat this as real learning or better bookkeeping. Hermes is betting that remembering your preferences, searching your past work, and reusing successful workflows is enough for an agent to feel smarter the next time you ask. (hermes-agent.nousresearch.com, hermes-agent.nousresearch.com)