Multi-agent (Graph) session support missing from AgentCoreMemorySessionManager

Start in src/bedrock_agentcore/memory/integrations/strands/session_manager.py and compare the existing single-agent methods with the three missing SessionRepository methods. Review repro_fixed.py and verify the proposed stateType and multiAgentId metadata approach against the append-only memory behavior. Done means a Graph can build and its serialized state can round-trip through AgentCoreMemorySessionManager without NotImplementedError.

Summary

AgentCoreMemorySessionManager (at
src/bedrock_agentcore/memory/integrations/strands/session_manager.py)
does not implement the three multi-agent methods on Strands'
SessionRepository interface (create_multi_agent,
read_multi_agent, update_multi_agent). As a result, any Strands
Graph that uses AgentCoreMemorySessionManager fails at graph
build time with NotImplementedError, before any AWS API call is
even attempted.

The single-agent methods (create_agent, read_agent,
update_agent, etc.) are fully implemented. The gap is only in the
multi-agent methods, which were added to SessionRepository
upstream and default to raising.

Impact

Any workflow that wants Graph-level session persistence with
AgentCore Memory is blocked. Examples:

• Classifier-routed multi-agent graphs (the Strands docs' own
  Interactive Customer Support
  canonical Graph example)
• Workflows that use Strands interrupts for human-in-the-loop
  steps, which require Graph nodes because the Agents-as-Tools
  pattern swallows interrupts

Reproduction

Minimal repro repo: https://github.com/akomandooru/agentcore-graph-repro/

Two scripts, broken and fixed, both using the real
AgentCoreMemorySessionManager from this SDK:

• repro_broken.py constructs AgentCoreMemorySessionManager,
  attaches it to a two-node GraphBuilder, and calls .build().
  Fails with NotImplementedError: MultiAgent is not implemented for this repository. No AWS credentials needed; the error fires before
  any API call.
• repro_fixed.py subclasses AgentCoreMemorySessionManager with the
  three methods and demonstrates graph state serialising and
  round-tripping cleanly.

File ".../strands/multiagent/graph.py", line 390, in build
    return Graph(nodes=self.nodes.copy(), ...)
File ".../strands/multiagent/graph.py", line 476, in __init__
    run_async(lambda: self.hooks.invoke_callbacks_async(MultiAgentInitializedEvent(self)))
File ".../strands/hooks/registry.py", line 306, in invoke_callbacks_async
    callback(event)
File ".../strands/session/session_manager.py", line 54, in <lambda>
    registry.add_callback(MultiAgentInitializedEvent, lambda event: self.initialize_multi_agent(event.source))
File ".../strands/session/repository_session_manager.py", line 332, in initialize_multi_agent
    self.session_repository.create_multi_agent(self.session_id, source, **kwargs)
File ".../strands/session/session_repository.py", line 58, in create_multi_agent
    raise NotImplementedError("MultiAgent is not implemented for this repository")
NotImplementedError: MultiAgent is not implemented for this repository

self.session_repository here is AgentCoreMemorySessionManager.
The MRO hops right past it into SessionRepository because the
AgentCore class doesn't override the multi-agent methods.

Proposed fix

Add three methods to AgentCoreMemorySessionManager, mirroring the
single-agent pattern. Full implementation in
repro_fixed.py. ~50 lines, mostly
boilerplate matching the existing single-agent methods.

Additions beyond the method bodies:

1. New stateType value: "MULTI_AGENT" (alongside "SESSION"
   and "AGENT").
2. New metadata key multiAgentId (alongside agentId) so
   multiple graphs in a session don't collide.

Payload is json.dumps(multi_agent.serialize_state()).
update_multi_agent delegates to create_multi_agent because
AgentCore Memory is append-only and list_events(max_results=1)
returns the latest by timestamp, same pattern as the existing
update_agent.

Willing to contribute

Happy to open a PR with the fix if the approach (new stateType
value, new multiAgentId metadata key, append-only with max_results=1
reads) is the one you'd take. Flagging as a question first in case
there's a planned design that differs.

Environment

• bedrock-agentcore 1.9.0 (latest published wheel as of
  May 11, 2026; verified that the three multi-agent methods are
  absent by inspecting the installed
  src/bedrock_agentcore/memory/integrations/strands/session_manager.py,
  which also matches main)
• strands-agents 1.39.0
• Python 3.13