DSH plugins for Memory & Context

Two problems live under this keyword. Context tools show you what the model is currently carrying: token composition, growth over a session, compaction and pruning events. Memory tools decide what persists: project facts, decisions, preferences, and recalled fragments injected at the start of a turn. Installing one does not give you the other.

Good memory is bounded and reviewable. Look for an explicit write policy, an audit surface you can read, and approval gates on anything the agent decides to remember on its own. Context tools earn their keep when they make compaction visible instead of silent.

How to evaluate

For context plugins: do they show composition (system, tools, history, injected files) and warn before a compaction drops what you still need? For memory plugins: where does data live (profile directory versus a remote service), how is recall scored, and can you delete or correct a stored fact without wiping the store?

Cross-agent and cross-session claims deserve extra scrutiny. A memory that “works everywhere” either centralizes your data or depends on a bridge — both are design choices you should be able to name before install.

What good looks like

A strong context tool is a dashboard and a log: tokens, layers, and a timeline of compaction. A strong memory plugin gates writes, stores source-backed facts rather than free-floating vibes, and injects a bounded slice at recall time so a long project does not drown in its own history. Graph-backed designs are useful when related facts surface together; flat embedding stores are fine when the corpus is small and the audit trail is clear.

The best setups in this directory run both: one plugin that explains the window, another that decides what survives after it closes.

Common failure modes

The usual failure is approval-free write paths that slowly fill the store with trivia. The second is unbounded recall that injects so much history the new turn starts already crowded. The third is confusing the two problems — a context meter does not remember last week’s decisions, and a memory plugin will not tell you why the current window is full.

Install order and boundaries

Measure the window before buying a store. Context and memory solve different failure modes; a single plugin claiming both must prove bounded retrieval and gated writes.