Rule-based context selection needs a pre-existing signal
Type: kb/types/note.md · Tags: context-engineering, agent-memory
A rule-based selector can load context for a specific case only when a rule-ready signal — a codified feature that already distinguishes that case from the other live candidates — is available to it. The signal can be a path, type, tag, tool argument, event payload, or workflow mode; it need not name the target, only separate it from the alternatives. The limit is structural: a rule cannot select on a distinction that none of its inputs encode. Codification makes such selection cheap, deterministic, and reviewable, but it cannot manufacture the distinguishing signal.
When no rule-ready signal is available yet, the system has three moves instead: wait for or ask for one, load a bounded superset so the case is present without being singled out, or infer relevance from the current task and candidate content — lexically or semantically. Inference can produce a targeted result without a predeclared route, but not without selection input: it derives a relevance relation from content instead of matching a distinction already encoded for the rule.
Push and pull do not change this. Pull supplies selection input directly through a query, path, or tool call. A path may itself be rule-ready; a natural-language query may instead supply content from which relevance is inferred. Push must obtain equivalent input upstream through a task identifier, current object, workflow mode, event payload, or input to a relevance classifier. A coarse trigger such as session start or a bare Write event can deliver generic context, but by itself cannot choose among multiple live candidates. Push names a delivery direction, not a way of knowing what to deliver.
The practical consequence inverts where the difficulty seems to lie. Storing a memory and injecting it are the easy parts; the hard part is the selector — deciding which sessions are the relevant ones — and that is exactly what needs the selection input. Having stored the curl lesson tells the system nothing about which sessions are about to run curl.
The bound is on selection, not presence: a bounded superset can hold the right item without any signal picking it out. Rule-based routing is reliable exactly when the rule-ready signal already exists; deriving relevance for a case still in flight belongs to inference rather than rule matching.
Relevant Notes
-
knowledge storage does not imply contextual activation — grounds: the activation gap this sharpens into a constraint on rule-based selection
-
agent statelessness means the context engine should inject context — contrasts: its open “how to identify what to inject” question; exact push needs an upstream signal
-
codification — defined-in: rule-based selection is the codified regime whose reach this bounds
-
frontloading spares execution context — contrasts: frontloading uses inputs known upstream of the call; this note characterizes when the required signal is upstream
-
agentic memory systems comparative review — evidenced-by: the cross-system push/pull split this mechanism explains
-
Activate Behavior-Changing Memory Before The Mistake — extends: applies the bound to pre-action activation and separates typed from inferred cues
-
Promotion selects for unreliable activation, and the regress ends only at an external trigger — extends: absent a rule-ready signal, develops the delivery fork — always-loaded slot or classifier as one recognition judgment relocated, not removed