aurora-serverless (Aurora Serverless v2) or rds-postgres (provisioned RDS). Use aurora-serverless for dev/staging and variable workloads.
enable_hindsight
bool
false
Optional Hindsight memory add-on (ECS Fargate + ALB for semantic + entity-graph retrieval). AgentCore managed memory with automatic per-turn retention is always on regardless of this setting — this flag only controls whether Hindsight is also deployed alongside it.
memory_engine
string
""
Deprecated. Use enable_hindsight instead. For backwards compatibility, setting this to "hindsight" is equivalent to enable_hindsight = true.
default_model_id
string
"anthropic.claude-3-5-sonnet-20241022-v2:0"
Bedrock model ID for the default Agent Template created on first deploy.
log_retention_days
number
30
CloudWatch log group retention in days. Set to 0 for never expire (not recommended for production).
ThinkWork always provisions an AWS Bedrock AgentCore Memory resource as
part of module.thinkwork. The agent container emits a CreateEvent for
every turn, and AgentCore’s built-in strategies (semantic, summarization,
user-preference, episodic) extract facts in the background. Retention is
automatic — you don’t need to configure anything or prompt the model to
call remember().
Variable
Type
Default
Description
agentcore_memory_id
string
""
Optional pre-existing AgentCore Memory resource ID to reuse. Leave empty to let the agentcore-memory module auto-provision one.
enable_hindsight
bool
false
Optional add-on: also deploy the Hindsight ECS service for semantic + entity-graph retrieval alongside managed memory.
hindsight_image_tag
string
"0.5.0"
Hindsight Docker image tag. Only used when enable_hindsight = true.
hindsight_retain_interval_hours
number
1
How often (hours) Hindsight runs the retain step. Only used when enable_hindsight = true.
hindsight_reflect_interval_hours
number
24
How often (hours) Hindsight runs the reflect/synthesis step.
hindsight_task_cpu
number
512
ECS Fargate CPU units for Hindsight tasks.
hindsight_task_memory
number
1024
ECS Fargate memory (MB) for Hindsight tasks.
When enable_hindsight = true, the agent container gains three additional
tools (hindsight_retain, hindsight_recall, hindsight_reflect)
alongside the always-on managed memory tools (remember, recall,
forget). The remember() tool dual-writes to both backends so explicit
memories land in both stores.
ThinkWork provisions an IAM service role that lets Bedrock Knowledge Bases read from the KB S3 bucket, invoke amazon.titan-embed-text-v2:0 for embeddings, and write vectors into Aurora Postgres via the RDS Data API. There is no OpenSearch Serverless collection — the vector index lives in the same Aurora cluster as the rest of your data (pgvector), so there is no separate vector-store cost line.
Knowledge Base configuration (chunking strategy, data source wiring, sync schedule) is managed at runtime through the Knowledge module, not through terraform variables.