Saltar al contenido principal

cache

Since v2.40.0 · Top-level declaration

Grammar​

cache <Name> {
    backend: redis | in_process          # optional — tier (default in_process)
    ttl: <duration>                       # optional — TTL; REQUIRED if non-pure
    key: [<param>, ...]                   # optional — key subset (default: all params)
    default: <true|false>                 # optional — auto-cover pure tools (max one)
    apply_to_effects: [<Effect>, ...]     # optional — effects covered (default [pure])
    invalidate_on: [<ChannelRef>, ...]    # optional — emit on any → flush this cache
}

cache declares a named, referenced result-memoization policy. It is the runtime dual of a proof the compiler already has: a tool whose effects: row is exactly pure is, by the language's own definition, safe to reuse for any input that repeats — so caching it is not a heuristic, it is a consequence of the purity proof (cacheability_is_a_type_not_a_convention).

A cache is referenced by a tool (tool.cache: <Name>) or a retrieve (retrieve <Store> { cache: <Name> }), exactly the way axonendpoint.cors: references a cors policy.

Surface​

cache is a top-level declaration.

type WeatherEvent { city: String }
channel WeatherUpdated { message: WeatherEvent }

# A project-wide default: every `pure` tool is cached, no per-tool
# annotation, because the type system already proved it safe.
cache DefaultPure { default: true }

# A widened, ttl-bounded, invalidation-wired cache for a NON-pure
# result (accepting bounded staleness).
cache WeatherCache {
    backend: redis
    ttl: 5m
    apply_to_effects: [pure, network]
    invalidate_on: [WeatherUpdated]
}

tool Fingerprint { provider: http  effects: <pure>  parameters: { input: String } }
tool Weather {
    provider: http
    effects: <network>
    parameters: { city: String }
    cache: WeatherCache
}

Fields​

backend: (optional)​

in_process (the default single-replica tier — a bounded LRU with per-entry TTL) or redis (the multi-replica tier). A closed catalog (axon-T866 on anything else). Which backend / how much capacity is a per-tenant, deployment-level knob; eligibility to cache is the source-level property.

ttl: (optional — but REQUIRED for a non-pure cache)​

A duration literal (10s, 5m, 1h). A pure-only cache may omit it (caching a proven-deterministic result forever is sound). A cache that widens apply_to_effects: beyond [pure] — or is used by a non-pure tool or any retrieve — must carry a finite ttl: (axon-T865): a non-deterministic result may not be cached forever. The TTL is precisely the "how stale am I willing to be" bound.

key: (optional)​

The subset of the covered tool's parameters: whose bound values form the cache key. Omitted ⇒ all bound parameters (the zero-friction default). Use it when one bound arg (e.g. a request_id) must not affect the key. The full key is content-addressed: a hash of (tenant ‖ cache ‖ tool ‖ tool-declaration-fingerprint ‖ output_type ‖ selected params), so a redeploy that changes the tool can never serve a stale result, and a tenant can never read another tenant's entry.

default: (optional)​

true ⇒ this cache auto-covers every eligible tool in the module without a per-tool cache:. At most one per module (axon-T863). A tool is eligible when its effects are a subset of the default's apply_to_effects (with [pure] as the implicit default set). A tool opts out with cache: none.

apply_to_effects: (optional)​

The effect classes this cache is willing to memoise. Omitted ⇒ [pure] (the only provably-safe default). Any member beyond pure is a widening (axon-W013 names each non-pure tool a default: true cache ends up covering) and forces a finite ttl:.

invalidate_on: (optional)​

Channel names (axon-T864 if undeclared). An emit on any listed channel flushes this cache's namespace — reusing the v1.6.0 pub/sub, not a second mechanism.

Budget interaction​

A cache hit short-circuits before the underlying tool/retrieve executes, so a budget { rate:/max: } quota (v2.28.0) is never decremented by a hit. Hits and misses are audited distinctly (cache:hit / cache:miss) from budget:consumed, so an adopter sees real cost savings.

What this primitive is NOT​

  • Not memory { decay: }. memory is semantic/conversational recall state (what the agent remembers about a conversation); cache is result-memoization of a deterministic or explicitly-accepted-stale tool/retrieve call. An agent can use both, for different reasons.
  • Not a way to cache a non-deterministic result forever. The compiler forbids it (axon-T865) — the same purity proof that authorises caching-forever is exactly what it withholds from a non-pure result.

See also​

  • axon://primitives/tool — declares the effects: pure proof this primitive derives cacheability from, and the cache: reference.
  • axon://primitives/channel — the invalidate_on: mechanism.
  • axon://primitives/memory — conversational recall state, a different layer.