Module injection 

Engine-native injection attention (selection policy for context injection). Engine-native injection attention: retrieval shaped for context injection rather than raw search.

Plain top-k recall is the wrong contract for injecting memories into a model’s context: the most similar memories to a conversation are echoes of that conversation, top-k always returns k items however weak the tail, and near-duplicates (a distilled fact and the turn it came from) co-rank. This module owns the selection policy that client hooks previously approximated with heuristics:

• Session provenance: memories originating from the querying session are never returned, they are in that context by construction.
• A working-memory ledger (exclude_ids) supplied by the client, holding what was already delivered within the current context lifetime.
• A relevance knee instead of a fixed floor: the candidate list is cut at the largest score gap, so weak tails vanish without a magic constant.
• Maximal Marginal Relevance over embeddings, so each selected item adds information the previous ones did not.
• Type quotas: distilled knowledge first, verbatim episodic capped, action notes never (they exist for distillation and session resume).
• User-pinned scope:always memories bypass every quality gate.
• Outcome learning: record_injection_outcome tracks shown vs used per memory; chronically ignored memories are demoted at selection time and used ones are reinforced.

Structs

InjectionCandidate

One injection-ready memory with its selection metadata.

InjectionConfig

Tunables for injection attention selection.

InjectionQuery

A request for injection-ready context.

Enums

SelectionReason

Why an item was selected, for introspection and client display.

Constants

ATTR_INJECTION_SHOWN

Attribute key: how many times this memory was injected into a context.

ATTR_INJECTION_USED

Attribute key: how many times an injected memory was drawn on by the reply that followed.