Per-agent memory & context registry
Each agent has its own memory (pending review, live entries, summary) and its own context registry rules. Pick an agent below to inspect and edit just that agent's state.
🧪 Test: queue a proposal
Append a fake learning to the queue for the selected agent. Watch the TowerRunTray (bottom-right) — within 15s a card will appear: cbdo: new DO proposal queued — "…". Then it shows up in the pending list below.
Pending review 0
Live entries 0
Rendered summary 0 chars
Recent events 0
Briefs 0
Each agent has one or more briefs — the system prompts the LLM receives for each task. Defaults ship in code; overrides live in KV and apply on the next cold-start (or immediately in the editing function). Clear an override to revert to the default. Builder briefs (those that interpolate runtime values like ${airContext}) accept overrides as templates — see the per-brief warnings.
Edit brief
View default (read-only)
Edit template
Upload reference documents (HTML, PDF, DOCX, PPTX). The LLM extracts
text + CSS digest, then derives a new partner-agnostic HTML template
with <!-- BLOCK:id --> markers. Save as a new template
under this category — leaves existing templates intact.
Context rules — tiering & weights
Tunes what THIS agent sees in the LLM context. Hot is always included. Warm competes by importance weight against a budget. Cold is vector-retrieved. Off excludes entirely. Persisted per-agent — applies on this agent's next refine.
What the numbers and colors mean
Tier (priority into the LLM context)
Bundle (visual grouping — color tag)
Weight (0–1000) drives ranking in the Warm tier (higher = more likely to fit in the budget) and serves as a tiebreaker for Hot sources if the hot total exceeds budget_tokens (lowest-weight hot dropped first). Cold sources are retrieved only if the query embedding hits them. Tuning happens at the leaf source (e.g. brandVoice), not at the bundle.