Aller au contenu

Agent Memory

Ce contenu n’est pas encore disponible dans votre langue.

An agentic worker can remember things across turns and across sessions. Memory has two tiers, configured in the optional memory block of the worker’s AgentConfig:

TierScopeBackingRunner requirement
Short-term (working memory)Within a conversationIn-memory providerBuilt in; needs a state store to run the worker at all
Long-termAcross conversationsChronicle (persistent)long-term-chronicle feature + GREENTIC_CHRONICLE_*
AgentConfig.memory
"memory": {
"short_term": { "enabled": true },
"long_term": { "provider": "chronicle", "credential_ref": "chronicle-main" }
}

Short-term memory is the worker’s scratchpad within a conversation. It is served by an always-on in-memory provider, so nothing extra is needed to build it — the model gets remember / recall behaviour when memory.short_term is enabled.

Long-term memory persists facts across conversations, backed by Chronicle (embedded SurrealDB by default, or an external graph store). Before each turn the worker can surface relevant long-term facts and inject them into the prompt.

Long-term memory is opt-in at the runner build: the runner must be built with the long-term-chronicle feature, and the tier stays silently disabled unless the Chronicle environment is configured. It never fails worker startup — an unconfigured tier is simply inactive.

VariablePurpose
GREENTIC_AW_REDIS_URLRequired on a server. Conversation state store; unset → dw.agent nodes disabled.
GREENTIC_CHRONICLE_*Long-term memory (Chronicle) connection. Absent → long-term tier inactive.