Add OCG-backed long-term agent memory (ports agentspan#298)

[NicholasDCole]

Ports the SDK side of the "OCG-backed agent memory" feature from agentspan-ai/agentspan#298 to the C# SDK. The server-side compiler that consumes the wire config lives in conductor-oss/conductor (agentspan-server: LongTermMemoryConfig.java, AgentConfig.longTermMemory/feedbackSink).

What was ported

• OCGMemoryStore (Conductor.AI/OCGMemory.cs) — a synchronous MemoryStore HTTP adapter over the OCG BFF:
  • Add → POST /api/v1/memories
  • Search → POST /api/v1/memories/search (folds the human good/bad signal into result content)
  • Delete → DELETE /api/v1/memories/{key}, ListAll → GET /api/v1/memories, Clear (fan-out)
  • GetFeedbackLinks → POST /api/v1/memories/{key}/feedback-links (mints signed good/bad capability URLs)
  • Auth via Authorization: Bearer <token>; the token is held client-side. Also exposes MemorySummary, FeedbackEvent, FeedbackLinks, and OCGMemory.BuildMemorySummarizer.
• Agent params (Conductor.AI/Agent.cs, + AgentBuilder fluent methods): SemanticMemory, MemorySummaryModel, FeedbackSink (Action<FeedbackEvent>).
• Serializer emission (Conductor.AI/AgentConfigSerializer.cs) — the most important piece. When an agent's SemanticMemory is backed by an OCGMemoryStore, emits:
  • longTermMemory: {ocgUrl, credential, agent, user, scope, maxResults, summaryModel} — credential is a server-resolvable secret NAME (OCG_PUBLIC_KEY), never the raw client token; summaryModel falls back to the agent's own model.
  • feedbackSink: {taskName: "{name}_feedback_sink"} when a feedback sink is set.
• Feedback-sink worker (Conductor.AI/WorkerManager.cs) — registers a {name}_feedback_sink worker (analogous to the existing callback/gate workers) so the compiled server path can hand the FeedbackEvent (distilled summary + signed good/bad URLs) to the user's FeedbackSink for out-of-band delivery. Best-effort. The capability URLs are never surfaced to the agent's LLM.
• Tests (Conductor.AI.Tests/OcgMemoryTests.cs, xUnit) — store adapter (add posts value not string_value/confidence; search folds good/bad signal; feedback-link route; non-2xx raises; blank url/agent rejected) and serializer emission (present / absent / summaryModel fallback / non-OCG store no-op).
• Example Conductor.AI.Examples/27_OcgMemory mirroring Python examples/118_ocg_memory.py.
• CHANGELOG entry.

What was skipped

Client-side pre/post-run memory hooks from Python runtime.py (_apply_memory_retrieval, _maybe_save_conversation_memory, _parse_summary_output) are not ported. This SDK has no client-side agent run loop — AgentRuntime.RunAsync is control-plane only (it serializes the agent, starts the workflow, and polls the server, which runs the LLM loop). Memory therefore activates entirely through the serialized longTermMemory config that the server-side compiler expands into pre-loop retrieval + post-loop distill/save/feedback. The corresponding runtime-hook unit tests were skipped for the same reason.

C#-specific adaptations

• OCGMemoryStore accepts an injectable HttpClient for testing (the C# analogue of Python's httpx.Client + MockTransport); the async HttpClient is driven synchronously to satisfy the sync MemoryStore interface.
• OCG-backed detection uses a type check (Store is OCGMemoryStore) rather than Python's duck-typed _base attribute probe.
• MemoryEntry is immutable (init-only), so Add returns the key without mutating the entry.
• FeedbackLinks is a typed record; MemorySummary/FeedbackEvent are plain classes.

Test results

dotnet test Conductor.AI.Tests — 65 passed, 0 failed, 0 skipped (9 new + 56 pre-existing). The main library and the new example both build clean.

🤖 Generated with Claude Code