Extend Memory Core with Explicit A2A Primitive

Context

This ticket was filed 2026-04-20 (session a38d25d1-3003-4d57-b0ac-0effe2c3507e) with a Chroma-metadata prescription: extend add_memory with a recipient field, surface get_inbox over the existing memory collection. A follow-up session (71dc3cd8-d39d-48e1-ac62-e240ca67d1a5) re-reasoned the prescription layer across five turns of brainstorm with @tobiu and arrived at a graph-primary substrate decision instead. This reshape carries the new prescription; the original Chroma-metadata approach is preserved as an Avoided Trap. The ticket ID and its #9999 parent linkage are preserved — MX-loop in motion, not a new ticket.

The Problem

Probabilistic semantic retrieval (query_raw_memories) cannot serve as an explicit A2A channel — proven empirically when cross-instance agents failed to hand off via add_memory convention. But the original fix (metadata on the memory collection) conflates two categories:

• Memories are open-shape, high-volume episodic content. Chroma substrate; semantic retrieval dominates access. Retention is accumulative.
• Messages are closed-contract, directed artifacts with invariants (sender, read-state, thread integrity). Relationship queries dominate access (inbox/outbox/thread/reply-chain). Lifecycle is explicit (read, archived, retracted).

Tagging memories with recipient works tactically for MVP but bakes category-conflation into the substrate. Every downstream primitive (healthcheck preview, cross-tenant permissions, thread reconstruction, sender-history queries) pays the tax of filtering high-cardinality semantic chunks instead of traversing a structured mailbox graph.

The Architectural Reality

• ai/daemons/DreamService.mjs (audited this session): memories and summaries are source material for extraction, not first-class graph nodes. The substrate exists but hasn't been used for mailbox-shaped work.
• SQLite (better-sqlite3) already backs the Native Edge Graph; FTS5 covers keyword search; Concept Ontology (#10030) supplies semantic edges.
• Existing graph tools (get_node, get_neighbors, search_nodes, query_hybrid_graph) absorb most mailbox access patterns. Mailbox-specific tools become thin wrappers, not a parallel tool tree.

The Fix

Mailbox as graph-primary directed messaging, built on Memory + Session + AgentIdentity as first-class graph nodes (prerequisites tracked in peer sub-epic + #10016 subs).

• Message node: {id, subject, bodyText, createdAt, readAt, archivedAt, priority, retracted}
• Edges: SENT_BY, SENT_TO (AgentIdentity | Role | Broadcast addressing), ORIGINATES_IN, RELATED_SESSION, REFERENCES_TICKET, IN_REPLY_TO, PART_OF_THREAD, TAGGED_CONCEPT (auto-emitted via #10030 integration).
• Optional Chroma mailbox collection for semantic body search — phase-gated, not v1.

This ticket is the parent sub-epic for mailbox work. Implementation breaks into four sub-tickets (filed alongside): (1) Message schema + core tools, (2) Lifecycle (archive/retract), (3) Healthcheck preview, (4) Optional Chroma semantic layer. Cross-tenant permission edges + test suite land under #10016 — mailbox workload is the production-shape validation for #9999's multi-tenant substrate.

Acceptance Criteria

• All four mailbox sub-tickets filed and linked
• Graph-first Memory artifacts sub-epic shipped as prerequisite (this ticket blockedBy it)
• AgentIdentity + OAuth2 work under #10016 shipped as prerequisite
• Mailbox tools consumed by at least one real agent handoff (A2A demonstration, not synthetic)
• Cross-tenant permission test suite (filed under #10016) passes with mailbox as workload
• (Post-merge) Agents adopt addMessage / listMessages for explicit handoffs; convention-based "leave a memory for the next session" is retired

Out of Scope

• Separate MCP server for A2A (original Avoided Trap; still rejected)
• Cross-physical-clone Chroma sync (remains local to the federated Memory Core)
• Human-facing mailbox UI

Avoided Traps

• Chroma-metadata prescription (original body of this ticket). Rejected after re-reasoning. Category conflation: messages and memories have different retention, invariants, and access patterns. Tagging memories with recipient works for MVP but bakes conflation into the substrate — every downstream primitive pays the tax. Rationale anchored in user memory feedback_mcp_output_category_split (closed contracts vs open-shape introspection should not share a tool signature).
• Separate A2A MCP server. Still rejected. Mailbox lives inside Memory Core MCP; extend the graph, don't fork the server.
• "Mailbox is just another kind of memory." Rejected. Messages carry explicit SENT_BY / SENT_TO / IN_REPLY_TO edges memories don't; conflating invents edge-shaped scalars and re-derives them at query time.
• Skipping memory-node lifting. Rejected. Messages need IN_REPLY_TO → Memory and REFERENCES_MEMORY → Memory edges. Memories must be nodes first — hence the peer sub-epic prerequisite.

Related

• Parent: #9999 (Cloud-Native Knowledge & Multi-Tenant Memory Core)
• Sibling (blocks this): Graph-first Memory artifacts sub-epic (filed this session)
• Sibling (blocks this): #10016 AgentIdentity + OAuth2 subs (filed this session)
• Context: Discussion #10137 (MX paradigm), #9535 (AI Village per-model account precedent)
• Heritage sessions: 5a521819-dc75-4549-888e-fcea818d0401 (MX discussion), a38d25d1-3003-4d57-b0ac-0effe2c3507e (original filing)

Origin Session ID: 71dc3cd8-d39d-48e1-ac62-e240ca67d1a5