Memory 6: Integrate Built-in Chat Deterministically

The current orchestrator owns an in-memory HistoryStore in src/server/app/runtime/common.py. It loads that history before execution and appends a successful user and assistant pair afterward in src/server/app/runtime/langgraph/chat.py.

Implement this integration in three increments.

Phase 1 Persistent Transcript

• Introduce a conversation-store protocol.
• Keep the current in-memory implementation as the disabled or fallback mode.
• Add an Agent Memory implementation using get_messages_async() and add_messages_async().
• Persist only after a completed response, preserving current failed-turn behavior.
• Ensure the streaming path writes exactly once after the final completion event.

Phase 2 Long-Term Retrieval

• Before each turn, search durable memory using exact user scope.
• Inject returned content in a clearly delimited, untrusted context block.
• Do not allow memory-derived text to authorize database writes, privilege changes, or other sensitive actions.
• Return non-content metadata such as result count and record types in ChatResponse and the SSE completion event.

Phase 3 Context Compaction

• Continue using raw messages for short conversations.
• At a configured token threshold, replace the older prefix with get_context_card_async().content plus a small recent-message tail.
• Do not pass both the complete transcript and a context card containing the same transcript.
• Oracle specifically recommends context cards when compaction should retain summary, topics, relevant durable records, and recent turns.

Reference: Use Agent Memory short-term APIs with LangGraph.

Recommended failure behavior:

• Explicit memory REST or MCP calls fail closed with a clear error.
• Chat retrieval may fail open and continue without memory, while reporting memory_status="unavailable".
• A failed post-response persistence call must not retract an answer already streamed to the client.
• Background extraction should initially be opt-in. Explicit "remember this" writes offer more predictable behavior for the first release.

Start with src/server/app/runtime/common.py and src/server/app/runtime/langgraph/chat.py to trace HistoryStore loading, successful-turn persistence, and streaming completion handling. Review the linked Oracle Agent Memory LangGraph guide before defining the conversation-store boundary. Done means the three phases cover durable retrieval and persistence, safe metadata and failure behavior, and threshold-based context-card compaction with exactly-once streaming writes.