Memory plugins have been the most experimental corner of the ecosystem, and this week they started looking like infrastructure. The designs that gained traction all share one property: you can see what the agent decided to remember. Opaque stores that only answer “it worked” are losing to gated writes and bounded retrieval — not because the demos got prettier, but because long projects finally hit the failure mode those demos never showed.
The approval gate becomes the norm
Memento continues to be the template. Writes are bounded, layered and gated behind approval, and a frozen snapshot is injected at recall time so results are reproducible. If you have been burned by a memory plugin that quietly accumulated noise over a long project, this is the design that fixes it.
The gate is not ceremony. Without it, a mis-summarised turn becomes permanent truth for every later session. With it, you trade a little friction for a store you can audit.
Mnemon grows its graph
Mnemon is now running across several agent runtimes from a single binary. Graph-based recall rather than flat embedding search shows up in the results: related facts surface together, and the supervision step keeps the graph from filling with trivia.
That supervision step is the feature that matters for directories to document. A memory graph without a filter is just a larger context window with worse latency.
Context is not memory
Worth repeating because the two get confused constantly. DSH Context tells you what is in the window right now and how it changed. Memory plugins decide what survives after the window closes. Installing one does not give you the other, and the strongest setups here have both.
Practical order: buy visibility first. When you can see compaction events, you know whether the problem is forgetting inside a session or amnesia across sessions. Only then add a store.
Smaller notes
At File shipped a faster workspace index, which matters in large repositories. Deep Whale added a variant with higher contrast panels. DSH Free Search had intermittent rate limiting from its public backend, a reminder to move to a keyed provider if search becomes load-bearing.
Rate limiting is not a scandal; it is the ordinary cost of a shared free resource. The mistake is treating the first throttle as a one-off outage.
Reading of the week
If you read one thing about agent memory, read the security guide section on the four permission surfaces. Memory plugins hold data about you by definition, and the place they hold it belongs in your evaluation alongside the retrieval quality.
What this means if you install
There is no single “must install” this week. If long projects are already losing context, buy visibility first (DSH Context) rather than another store. If the failure mode is “nothing survives next week’s session,” then add gated memory. Treat Memento and Mnemon as answers to different questions, not competing SKUs on one shelf. When a free search backend becomes load-bearing, plan the keyed path before the first rate limit — throttling is the normal cost of a shared resource, not a one-off outage.
Pin versions while snapshot schemas settle. And after any multi-plugin install, check the registered tools: two memory plugins that both claim the “memory” capability can override each other in ways neither README admits.