strands-memorysync

MemorySync for Strands Agents (AWS's agent SDK): automatic memory injection before every model call, duplicate-proof turn-by-turn extraction, and a crash-proof lifecycle — with zero framework dependencies.

pip install strands-memorysync

Quick start

from strands import Agent
from strands.memory import MemoryManager
from strands_memorysync import MemorySyncStore

store = MemorySyncStore(
    api_key="ms_...",          # or MEMORYSYNC_API_KEY
    user_id="customer-42",     # required — who these memories belong to
    session_id="support",      # tags the facts from this conversation
)

agent = Agent(memory_manager=MemoryManager(stores=[store]))
agent("which seat should I book?")

The MemoryManager consults the store before every model call (automatic injection — not tool-gated), and the extraction sink sends the user's messages to MemorySync, which extracts the durable facts they contain. The agent's replies are not sent: assistant replies are not stored as memories.

Why this one

Semantics worth knowing

• User messages are sent to fact extraction; only durable facts are stored. Each user message in a batch goes to MemorySync as plain text with role: "user" under the strands::<session> session scope. The server extracts the durable facts it contains and stores only those, tagged with session_id: "strands::<session>"; filler such as "ok" or "thanks" yields none, and the message text itself is not stored. Deterministic seeds let the server recognise manager retries and batch overlaps as the same turn and not extract it twice.

• Assistant replies are not sent (they are not stored as memories); tool and system messages never are. A batch with nothing the user said makes no request at all.

• Turn-by-turn extraction. With the default extraction=True, Strands extracts every 5 invocations (the synchronous agent(...) entry point also flushes after each call; invoke_async does not). For turn-by-turn extraction on every path, pass an invocation trigger:

  from strands.memory import InvocationTrigger
  
  store = MemorySyncStore(user_id="customer-42",
                          extraction={"trigger": InvocationTrigger()})
  

  MemorySync's writes are cheap and idempotent, so this is safe, and memories are recallable from the very first exchange.

• AddMessagesContext.sequence_numbers travel as each user turn's sequence metadata, which its facts carry.

• Free-tier quota exhaustion is silent by design (adds accepted-without- storing, reads empty); evaluation keys surface strict 429s.

• store.get_tools() returns memorysync_search + memorysync_save (@tool-decorated) — enabled via the MemoryManager's tool config.

Configuration

Development

pip install -e . strands-agents pytest pytest-asyncio "httpx>=0.25,<1"
python -m pytest tests -q     # 26 tests incl. a REAL Agent turn (stub Model)

License

MIT