Curate core memory (own your working memory)

Core memory is the small, always-in-context block of the highest-signal facts about the current project — conventions, gotchas, the architecture one-liner. It is size-bounded (it costs context on every turn), so it must be curated, not just accumulated. This is MemGPT's insight: the agent deliberately owns its working memory, distinct from the auto-accumulated KG triples and RAG chunks.

It is wired to Hermes's native MEMORY.md, so anything you put here is auto-loaded into context — one source of truth, no parallel store.

At task START

• core_memory_get() — read the block to orient. Trust it, but verify any file/symbol it names still exists before acting (it reflects a past moment).

At task END / boundaries (the curation pass)

1. core_memory_get() and ask: what's the highest-signal thing learned this task that a future session must know (a convention, a non-obvious gotcha, a key decision)? Not routine facts — those belong in the KG / RAG.
2. Add it: core_memory_append("<one crisp fact>"). If it's rejected for overflow, that's the signal to prune.
3. Prune / update: core_memory_replace(old=..., new=...) to fix a stale fact, or core_memory_replace(block="<the whole rewritten block>") to do a full curation pass — keep the few highest-signal facts, evict what's stale or now-obvious. Keep it well under the limit; tighter is better.

Rules

• High-signal only. If it's derivable from the code or already in the KG, it doesn't belong in the always-in-context block.
• Bounded. The append tool rejects overflow on purpose — respond by pruning, never by trying to grow the window.
• Deliberate. This is a conscious edit, not a dump. One good line beats five vague ones.
• This curation pass is itself a future GEPA optimization target (it improves with feedback over time).