DynamoDB

(serverless KV + session state + agent memory)

Mental model (when it wins)

• Single-digit ms KV at any scale, serverless scaling.
• Best for: session state, agent run state, tool execution ledger, idempotency keys, rate-limit counters, feature flags, metadata.

The knobs that matter

• Key design (most important): partition key + sort key shape your scalability.

  • Enemy: hot partitions (one key gets most traffic).

• Capacity mode

  • On-demand: default for unpredictable traffic.
  • Provisioned + autoscaling: cheaper for steady load.

• GSIs: powerful, but each GSI is effectively “another table” in cost + write amplification.

• TTL: auto-expire session/memory items; TTL deletes don’t consume WCUs. ([AWS Documentation][2])

• Streams: for CDC into pipelines (lambda/eventbridge/kinesis).

• PITR: point-in-time recovery (prod default for critical tables).

• Table class: Standard vs Standard-IA when storage dominates (IA cheaper storage, slightly higher request pricing). ([Amazon Web Services, Inc.][1])

Pricing mental model (back-of-envelope)

From AWS example (US East):

• Writes: ~$0.6 per 1M writes
• Reads: ~$0.1 per 1M reads
• Storage: ~$0.25 per GB-month ⇒ ~$25 per 100GB-month (or ~$250 per TB-month) ([Amazon Web Services, Inc.][1]) And Standard-IA flips that: storage much cheaper, requests slightly more expensive. ([Amazon Web Services, Inc.][1])

Senior heuristic: DynamoDB bills are usually either:

1. Storage-dominated (big tables) → consider Standard-IA + TTL, or
2. Write-amplification dominated (many GSIs / item updates) → reduce indexes or shrink item size.

Common “agent memory” pattern

• Table: PK = user_id, SK = conversation_id#timestamp (or run_id#step)
• TTL on each item (e.g., 7–30 days)
• Optional GSI for run_id lookups

Terraform template (on-demand table + TTL + PITR + optional stream)

resource "aws_dynamodb_table" "mem" {
  name         = var.table_name
  billing_mode = "PAY_PER_REQUEST" # on-demand
  hash_key     = "pk"
  range_key    = "sk"

  attribute { name = "pk"; type = "S" }
  attribute { name = "sk"; type = "S" }

  point_in_time_recovery { enabled = true }

  ttl {
    attribute_name = "expires_at"  # epoch seconds
    enabled        = true
  }

  server_side_encryption { enabled = true }

  # Optional: change data capture
  # stream_enabled   = true
  # stream_view_type = "NEW_AND_OLD_IMAGES"

  tags = var.tags
}

variable "table_name" { type = string }
variable "tags"       { type = map(string) default = {} }

---