Itemized Bullet Context
Represent an evolving agent context as a set of small, individually addressable bullets with IDs and helpful/harmful counters, not one monolithic prompt, so updates and retrieval can be localized.
An itemized bullet context stores accumulated knowledge — reusable strategies, domain concepts, common failure modes — as a collection of structured bullets rather than a single prose system prompt. It is the core representational choice of ACE's generator–reflector–curator loop.
Each bullet has two parts:
- Metadata — a unique identifier plus counters recording how often the bullet was marked helpful or harmful when the agent used it.
- Content — one small unit of knowledge (a tactic, a schema fact, a pitfall).
This is close to a memory entry in agent-memory systems (compare Zettelkasten agent memory notes), but the usage counters turn the store into something the agent's own task outcomes can grade: while solving a task the Generator flags which bullets helped or misled, and that signal steers the next corrective update.
In ACE's published prompts a bullet renders as
[ctx-00263] helpful=1 harmful=0 :: <content>, and bullets are grouped under
named sections such as strategies_and_hard_rules,
apis_to_use_for_specific_information, verification_checklist, and
formulas_and_calculations (the first three appear in the AppWorld prompts,
the last in the FINER prompts). IDs and counters are managed by system code
and kept out of model-authored content (see the
curator delta operation contract).
How usage is attributed differs by prompt variant. In the FINER prompts the
Generator outputs the IDs it used (bullet_ids), and the
Reflector returns a bullet_tags
list (helpful, harmful, or neutral per ID). In the AppWorld prompts the
Reflector is told to tag the bullets, but neither listed output schema has a
bullet field.
Why itemize instead of keeping one monolithic prompt:
- Localization — an update touches only the relevant bullets; unrelated knowledge is not re-emitted and so cannot be silently dropped (the mechanism behind context collapse).
- Fine-grained retrieval — the agent can attend to the most pertinent bullets rather than an undifferentiated block.
- Incremental adaptation — merging, pruning, and de-duplication become per-item operations (append, update by ID, embedding comparison) instead of regeneration, which is what makes incremental delta updates and grow-and-refine maintenance possible.
Design implication: if a harness lets a model edit its own instructions or memory, give the editable surface addressable units (IDs) and per-unit provenance/usage signals; whole-document rewrites forfeit all three properties.
Itemization also enables selective unlearning. A learned context is human-readable, and each lesson is a separate bullet, so a specific item can be deleted when it turns out outdated, wrong, or subject to an erasure obligation (for example GDPR right to erasure). Weight updates offer no such surgical removal.
Sources
- ACE (Zhang et al.), §3.1, §5, App. F