A tool called Stato starts from a specific complaint: an agent session is stateless, so whatever it figured out — which migration script is safe to rerun, which test actually covers the retry path, which config flag nobody documented — disappears when the session ends, unless someone happens to write it down somewhere the next session will read.
Stato's answer is to make that writing-down a deliberate, portable action. stato crystallize packages what an agent learned into a snapshot. From there you can slice out one specific skill, graft an external one onto a different project, or merge expertise gathered from more than one source. The explicit pitch is cross-platform: plan in one tool, build in another, carry the snapshot with you rather than starting over.
A design fork, not just a new format
The useful comparison here isn't a feature list, it's a fork in how "don't make the agent re-learn this" gets solved at all. A tool called Hivemind takes the other path: a shared, self-hosted store that automatically mines patterns from live sessions across a team's agents and propagates them to every connected agent in real time, with an optional approval step before a pattern spreads. Continuous and automatic versus explicit and curated — the same underlying problem, opposite defaults about who decides what gets kept.
I don't think one of these is simply correct. An automatic system catches what nobody thought to write down. A curated snapshot can't silently promote a fluke from one session into a rule everyone inherits. The second property matters more to me, for the same reason I keep coming back to it in A retro only counts if you can measure it: a finding that sounds right after one session is still a guess until something checks it again. Automatic propagation makes that check easier to skip, not harder.
AGENTS.md sits next to both of these without competing with either — it describes what a project is (how to build it, what the conventions are), not what an agent learned doing the work. Useful context, different problem.
The same checkpoint question, now portable
What to keep before an agent forgets argues that a good checkpoint preserves the goal, the exact evidence, and the rejected assumptions — not just a plausible summary. Stato's crystallize step is the same obligation with a new constraint: the snapshot now has to survive leaving the tool that produced it. A summary that only makes sense inside one agent's internal state doesn't travel. Evidence and a stated boundary do.
That portability is the part I'd actually test before trusting it on anything real: not whether the snapshot format is clean, but whether a skill grafted from one project onto an unrelated one carries its original boundary conditions with it, or just the parts that sounded generalizable at export time.
What I'm taking from it
The fork itself, more than either tool. Before adopting a shared-memory or portable-expertise system, I'd want a straight answer to one question: when a pattern moves from one place to another, did a person decide that, or did the system decide it on the pattern's behalf? Automatic is not free — it trades a visible decision for a convenient one.
Sources and further reading
Stato, shared on Show HN. Hivemind, the automatic-propagation comparison; its reported cost and token-reduction figures are the project's own numbers, not independently checked by me. I have not run either tool myself.