Why a curator role

The curator is one of the harness’s load-bearing roles. It’s the only role allowed to delete memory. Without it, raw.md grows linearly with iterations and no other role would have the bandwidth to read it.

What the curator does (in order)

1. Process pending issues (pending-issues.jsonl → issues.json). Validators emit findings in JSONL; curator dedupes by code-pointer + title, increments attempt counters on duplicates, auto-closes stale issues.
2. Promote raw observations. Workers/validators/etc append to raw.md as they go. Curator promotes recurring patterns to MEMORY.md (if mission-specific) or identity/<role>.md (if generalisable across missions).
3. Update summary.md. Bounded at 40 lines. Re-derived from MEMORY.md every run. This is what every role sees auto-injected.
4. Truncate raw.md. Keep last 10 entries as recent-history ballast.
5. Maintain identity files. ~/autonomous-harness/identity/<role>.md per role. Append-only. Compaction trigger: file exceeds 60 lines.

Why “compaction” is a separate role

Other frameworks bundle memory maintenance into every agent. The result is:

• Workers spend tokens deciding what’s worth remembering.
• Different workers disagree on what’s important.
• Mid-mission deletes (because two workers thought the same lesson was wrong) cause memory churn.

Splitting compaction into a dedicated role gives:

• One place to review the “what should we remember” prompt.
• Atomic mutations (curator runs, all roles see the new memory next iteration).
• Auditable history — raw.md is the working set, MEMORY.md is what the curator distilled.

When the curator fires

• On terminal events: DONE and ESCALATE. Always.
• On manual trigger via UI.
• (Optional) After every N iterations — not currently auto-fired, but easy to add.

The “TRICK:” promotion path

Workers can flag observations they think are durable:

[2026-04-24T10:05:00Z] worker F-023 TRICK: psql -f drizzle/*.sql applies migrations in filename order; do NOT use npm run prisma:migrate

Curator sees this and:

1. Counts how many distinct workers (by feature id) have flagged something with the same TRICK: lemma.
2. If ≥ 2: promote to identity/worker.md under “Patterns I’ve learned”.
3. If just 1 but it references codebase-specific paths: promote to this mission’s MEMORY.md Lessons.

The worker doesn’t decide where the observation belongs. They just flag. Curator decides.

Why this design wins

• Single point of truth for what “memory” looks like.
• No coordination between roles for memory writes; everyone appends to raw.md, curator handles the rest.
• Bounded growth. 40-line summary, 200-line MEMORY, 60-line identity-per-role. The system can’t accidentally ingest more than it can re-read.