Skip to content

Multi-Agent Team Workflow

For teams with multiple developers, each with their own AI agents.

Daily Workflow

  1. Morning: Each agent starts fresh and loads the consolidated central store.
  2. During the day: Each agent writes to their own local memory store (fast, no contention). All agents also read from the central store for recall.
  3. End of day: Each agent's local store is merged into the central store via POST /merge or POST /end_of_day.
  4. Overnight: Nightly consolidation runs -- clustering episodic memories into semantic memories, deduplication, pruning, and contradiction detection.
  5. Next morning: Agents load the freshly consolidated central store.

Setup

# Start the central server (accessible to all agents on the network)
python -m integrations.claude-code.memory_server \
    --host 0.0.0.0 --port 7832 \
    --data-dir /shared/team-memory \
    --auto-save 300

# Each agent includes their agent_id in requests
curl -X POST http://memory-server:7832/ingest_and_recall \
    -d '{"agent_id": "alice-agent-1", "session_id": "sprint-14", "role": "user", "content": "..."}'

# End of day: run the full nightly workflow
curl -X POST http://memory-server:7832/end_of_day -d '{}'

Contradiction Detection

When agents submit conflicting information, the system flags it with agent attribution:

curl -X POST http://memory-server:7832/contradictions -d '{"threshold": 0.7}'
# Returns: who said what, so the team can investigate

Scope Visibility

Memories have scopes that control visibility:

  • TEAM: visible to all agents (default)
  • SHARED: visible to all agents
  • PRIVATE: visible only to the creating agent