Skip to main content

context

Since v0.1.0 · Top-level declaration

Grammar​

context <Name> {
    memory: <ephemeral|none|persistent|session>   # optional — conversational memory scope
    language: "<bcp47>"                           # optional — preferred output language tag
    depth: <shallow|standard|deep|exhaustive>     # optional — reasoning thoroughness
    max_tokens: <integer>                         # optional — output budget
    temperature: <0.0..2.0>                       # optional — sampling temperature
    cite_sources: <true|false>                    # optional — force citations on every claim
    now: "<IANA-tz>" # optional (v2.46.0) — the frame's declared cognitive timezone
}

context declares the conversational frame a flow operates within: how much memory it retains across turns, how deep it reasons, how many tokens it can emit, what temperature governs sampling, and whether citations are mandatory. A flow that runs without an explicit context inherits the runtime's defaults; a flow that runs within a declared context binds those defaults explicitly + auditable.

This is the second primitive an agent touches — after persona (who the agent is), context says how the agent deploys that identity in a given conversation.

Surface​

context is a top-level declaration. It is not nested inside a flow, a persona, or a daemon. A run statement binds it via within <Context>.

context LegalReview {
    memory: session
    language: "en"
    depth: exhaustive
    max_tokens: 4096
    temperature: 0.3
    cite_sources: true
}

Fields​

memory: (optional)​

A single identifier from the closed memory-scope catalogue:

ValueMeaning
ephemeralNo retention. Each turn starts fresh.
noneAlias for ephemeral. Explicit "intentionally amnesiac".
sessionRetains across the conversation; cleared on disconnect.
persistentRetains across reconnects, scoped by tenant + user.

The type checker rejects unknown values (axon-frontend::type_checker::VALID_MEMORY_SCOPES).

language: (optional)​

A string literal containing a BCP-47 tag ("en", "es-CO", "pt-BR", …). Sets the preferred output language for the flow. Overrides the request locale + the persona's language: field.

depth: (optional)​

A single identifier from the closed depth catalogue: shallow | standard | deep | exhaustive. The runtime maps this to backend-specific reasoning effort (max-thought tokens, chain-of-thought budget, retrieval breadth).

max_tokens: (optional)​

A non-negative integer literal. Hard cap on the flow's output length. Independent of the persona's confidence_threshold: — max_tokens bounds length, confidence_threshold bounds uncertainty.

temperature: (optional)​

A numeric literal in [0.0, 2.0]. Sampling temperature forwarded to the backend. 0.0 requests greedy decoding; typical values are 0.0–0.4 for analytical work, 0.7–1.0 for creative work. Above 1.0 the runtime emits an axon-W007 warning (sampling becomes near-uniform).

cite_sources: (optional)​

A boolean. When true, every factual claim emitted while the context is active MUST carry an [evidence: ...] citation — identical semantics to the persona-level field of the same name. The two compose: if either the persona OR the active context demands citations, citations are required.

now: (optional, v2.46.0)​

An IANA timezone string literal ("America/Bogota", "UTC") — the frame's declared cognitive timezone. Every step running within this context carries the run's single captured instant, rendered in this zone, in its system prompt — unless the step declares its own now: (step overrides frame). This is time_is_an_explicit_input (v2.27.0) applied to cognition: a scheduling agent must say WHOSE time it reasons in; "the server's local time" is the ambient read this doctrine forbids. Format-checked at compile time (axon-T892); the envelope records (captured_utc, tzdb_version, zones) for byte-exact replay. Absent ⇒ no temporal injection (byte-identical to pre-v2.46.0).

Runtime behaviour​

At deploy time, every context declaration lowers to a ContextDefinition IR node. At run-time the binding happens once:

run AnalyzeContract(doc)
    as LegalExpert            # persona — who
    within LegalReview        # context — how (this declaration)
    constrained_by [NoHallucination]

The context's fields are injected into:

  • The system prompt's frame section (depth, language).
  • The backend's sampling parameters (temperature, max_tokens).
  • The shield + anchor layer (cite_sources).
  • The session-management layer (memory:).

Unlike persona, a context can be swapped per-run without re-declaring the persona — useful when the same expert needs to operate in different conversational frames (one-shot Q&A vs. multi-turn dialogue).

What this primitive is NOT​

  • Not a system prompt. A context is structured metadata. The system prompt is one of its lowerings, not its definition.
  • Not nested inside a flow. A flow references a context by name on the run statement; it does not declare one inline.
  • Not a persona. Persona = identity (who); context = frame (how). The two compose; they are not interchangeable.
  • Not a session. memory: session IS the conversational memory scope; session (the primitive) is a v2.3.0 duality-typed dialogue protocol — different concept.

See also​

  • axon://primitives/persona — the who that pairs with this how.
  • axon://primitives/run — the binding site (within <Context>).
  • axon://primitives/memory — durable, addressable memory stores; complements the conversational memory: scope here.
  • axon://primitives/flow — the orchestration that consumes the context.