Memory

Memory is durable tenant + repo knowledge stored under memory/**/*.md (plus equivalent paths inside .coro/memory/ for repo-local notes). It captures operational facts the next job should not have to rediscover—build quirks, incident learnings, naming preferences.

On-demand reads

Agents call read_memory({ file: "…" }) (and related listing helpers) when they need a slice of context. The harness does not paste the entire tree into every prompt by default, which keeps runs fast and avoids stale global narrative dominating fresh repo facts.

Author memory as small, addressable files with stable names so agents can target them predictably.

Merge semantics

Across layers, memory files append with provenance banners—new sections never silently delete prior ones. That makes memory safe for multi-tenant evolution: everyone adds signal, nobody “wins” by obliterating history in merge.

Subtractive grooming

Because append-only growth is monotonic, Coro provides the memory-curator workflow executed by its namesake agent—the only role allowed to collapse, deduplicate, or delete memory while preserving informational content.

Curator output still lands in Git through the same human-reviewed PR pipeline as other intelligence updates; it is not a silent database mutation.

Cross-link: Self-improvement for how proposals reach Git, and Agents for how the curator fits the roster.