pydantic-ai-dynamodb-memory

An Amazon DynamoDB MemoryStore for the PydanticAI harness Memory capability — the per-user Markdown notebook the harness injects before every model call and lets the model write through write_memory, on a shared, durable table.

pip install pydantic-ai-dynamodb-memory

from dataclasses import dataclass
from pydantic_ai import Agent
from pydantic_ai_harness.memory import Memory
from pydantic_ai_dynamodb_memory import DynamoDBMemoryStore

@dataclass
class Deps:
    user_id: str

store = DynamoDBMemoryStore(table_name="agent-memory")          # table auto-created (PAY_PER_REQUEST)
agent = Agent("anthropic:claude-sonnet-5", deps_type=Deps,
              capabilities=[Memory(store=store, namespace=lambda ctx: ctx.deps.user_id)])

agent.run_sync("I moved to Hanoi and hate early flights.", deps=Deps("kamal"))   # model calls write_memory
agent.run_sync("Book me something to Bangkok.", deps=Deps("kamal"))              # MEMORY.md injected first

How it works

• One table, pk = f#<path> for memory files (content, version, operation_id) and o#<operation id> for operation receipts. Auto-created; pass create_table=False to manage it yourself.
• Compare-and-swap is a DynamoDB conditional write: create requires attribute_not_exists(pk), replace and delete require version = :expected. A stale version raises the harness MemoryConflictError, never an overwrite.
• Operation receipts (MemoryOperation) are reserved with a conditional put, completed after the mutation, and dropped if the mutation fails, so durable-execution replays return the original result and a reused id with different arguments raises MemoryOperationConflictError.
• search is the harness's bounded lexical search over the scope's files (SearchableMemoryStore); list_paths is a prefix scan, sized for notebook volumes.
• Async protocol methods run the boto3 calls in a worker thread.

Writing to memory from outside the model, e.g. from the agentstate-reducer on_prune hook: await append_memory(store, f"{user_id}/main/facts.md", text) from pydantic-ai-memory-core does a CAS-safe append with retries.

Permissions: DescribeTable, CreateTable, GetItem, PutItem, UpdateItem, DeleteItem, Scan.

Docs: https://skamalj.github.io/agentstate-reducer/ · part of pydantic-ai-memory

License

MIT