Nick Trenkler
hermes agent and claude-mem retain their positions in the weekly ranking for the third time in a row, with 22.8% and 15.8% growth respectively
here are the highest-starred github repos of the week:
andrej-karpathy-skills by forrest chang +40,732
hermes agent by nous research +25,081
claude-mem by thedotmack +10,356
voicebox by jamie pine +5,198
evolver by evomap +4,376
genericagent by lsdefine +4,223
agentic ai and persistent memory aren’t slowing down - yet the picture seems far more versatile than the last week. excited to find out where this goes!
the six repos and what they do
each repo on the list solves a different problem for people who build with ai agents. here is what they do, in ranking order:
- andrej-karpathy-skills by forrest chang – a single 65-line
CLAUDE.mdfile for claude code. it turns andrej karpathy's post on llm coding pitfalls into four rules: think before coding, simplicity first, surgical changes, goal-driven execution - hermes agent by nous research – an open-source agent built to grow with its user. v0.9.0 shipped on april 13 with android support, imessage, wechat and a local web dashboard. v0.10.0 followed on april 16
- claude-mem – records what an agent does in a session, compresses it with ai and feeds the relevant context into the next session
- voicebox by jamie pine – an open-source ai voice studio for voice cloning and dictation
- evolver by evomap – a self-evolving engine for ai agents that logs every change as genes, capsules and events
- genericagent – an agent that grows its own skill tree from a 3.3k-line seed; its authors claim 6x lower token use
five of the six are tools for the agents themselves, not apps built on top of them. voicebox is the only end-user product on the list.
why agent memory keeps winning
karpathy-skills added 40,732 stars in a week. that is more than hermes agent and claude-mem gained together.
hermes agent and claude-mem have now held their spots for three weeks in a row. claude-mem targets a known weakness: an agent forgets what happened in the last session. it stores a compressed history of past work and brings it back when it is relevant.
evolver and genericagent go one step further: the agent changes its own skills over time. two such projects in one weekly top show where builders are looking.
update, october 15, 2026: where the six repos are now
six months later, all six repos are still active on github. the table shows their stars and releases as of october 15.
| repo | stars on oct 15 | releases since apr 23 | latest release |
|---|---|---|---|
| hermes agent | 254,085 | 32 | v0.22.2, oct 15 |
| andrej-karpathy-skills | 220,066 | 0 | no releases, last commit apr 20 |
| claude-mem | 101,006 | 100+ | v13.39.0, oct 13 |
| voicebox | 57,586 | 1 | v0.5.0, apr 25 |
| genericagent | 14,318 | 6 | desktop v0.2.0, aug 24 |
| evolver | 8,687 | 100+ | v2.0.42, sep 30 |
hermes agent now has the most stars of the six. in the april week karpathy-skills gained more, but today hermes leads by 34k.
karpathy-skills has not changed since april 20. its CLAUDE.md is still the file committed on january 27, and the repo now lives under the multica-ai organization.
the two repos that held their spots for three weeks kept shipping. hermes agent went from v0.10.0 to v0.22.2. claude-mem published more than 100 releases and is at v13.39.0.
evolver is the smallest of the six today with 8,687 stars. about half of them came in the single april week of this ranking.
the agent tools on this list kept shipping after april. the one-file repo kept collecting stars without a single new commit.
sources
- andrej karpathy's post on llm coding pitfalls – january 2026, the source of the rules in
CLAUDE.md - hermes agent v0.9.0 release notes – april 13, 2026
- hermes agent releases – v0.10.0 (apr 16) to v0.22.2 (oct 15)
- claude-mem releases – v13.39.0, oct 13
- evolver releases – v2.0.42, sep 30
- genericagent releases – desktop v0.2.0, aug 24
