Cross-Agent Memory

Multiple agents owned by the same user can share memories within a single tenant. This enables collaborative AI workflows where context transfers naturally between agents.

Memory Isolation Model

Synapse Layer enforces a strict isolation model:

Tenant = User

Each authenticated user maps to exactly one tenant. A tenant is the outermost boundary for memory access. Cross-tenant memory access is impossible — there is no API to query another user’s memories.

Agent = Namespace

Within a tenant, each agent_id acts as a namespace — a logical grouping of memories. Agents within the same tenant can read each other’s memories by performing recall without an agent_id filter.

How Cross-Agent Recall Works

By default, recall queries are scoped to the agent that stored the memory. To access memories from other agents within the same tenant, omit theagent_idfilter during recall:

python
# Agent-scoped recall (default)
results = await memory.recall("user preferences", limit=5)
# Returns only memories stored by this agent

# Cross-agent recall (tenant-wide)
results = await memory.recall(
    "user preferences",
    limit=5,
    agent_id=None  # omit agent_id filter
)
# Returns memories from ALL agents in this tenant

The recalled memories include the agent_id of the originating agent, allowing the caller to distinguish between its own memories and those from collaborating agents.

The agent_id Namespace

The agent_id serves as a logical namespace:

  • Required at store time — every memory must be tagged with an agent_id
  • Optional at recall time — omitting it enables cross-agent recall
  • Free-form string — any value is accepted (e.g., "coding-assistant", "research-agent")
  • Default value — if not specified, defaults to "default-agent"

Secure Agent Handover

For structured context transfer between agents, Synapse Layer provides Secure Agent Handover (currently in beta). Unlike passive cross-agent recall, handover is an active process:

  • Creates a cryptographic session with SHA-256 token
  • Packages selected context with metadata
  • Enforces one-time consumption and TTL expiry
  • Available via the neural_handover MCP tool

Note

Secure Agent Handover is in beta. It is available on all plans during the beta period. Enterprise plans receive higher limits and priority SLA.

Boundaries

Hard Boundaries (Cannot be Changed)

  • Cross-tenant memory access is impossible
  • No admin or superuser override exists for tenant isolation
  • Tenant ID is derived from authentication, not from request parameters

Important Notes

  • Cross-agent recall returns all memories in the tenant, not just from specific agents
  • There is no per-agent access control within a tenant — all agents share the same memory pool
  • Memory ownership is tracked by agent_id but does not restrict read access within the tenant