feat(memory): OCG-backed agent memory with auto-summarized save + hum… - #298
Open
NicholasDCole wants to merge 5 commits into
Open
feat(memory): OCG-backed agent memory with auto-summarized save + hum…#298NicholasDCole wants to merge 5 commits into
NicholasDCole wants to merge 5 commits into
Conversation
…an feedback links Add OCGMemoryStore (a synchronous MemoryStore HTTP adapter over the OCG BFF) and wire long-term memory into the agent runtime: - Pre-run: relevant agent/user-scoped memories are retrieved and injected into a copy of the agent's instructions (never mutating the shared agent). - Post-run: the conversation is distilled by an internal summarizer sub-agent (structured MemorySummary output; reuses the agent's model unless memory_summary_model is set) and saved as a 'conversation:<session>' memory. Best-effort and recursion-guarded - never fails the primary run. - Feedback is human-only: the runtime hands a FeedbackEvent (distilled summary + signed good/bad capability URLs) to the agent's feedback_sink for out-of-band delivery (e.g. a Zendesk ticket). The URLs are never shown to the agent's LLM, so agents can only create and read memories. Adds Agent params semantic_memory/memory_summary_model/feedback_sink, the OCGMemoryStore.feedback_links() mint call, example 118, and unit tests. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…path Extend OCG-backed agent memory beyond the client-side run() wrapper so it also works on the compiled (deployed/webhook) execution path. Python SDK: - Serialize semantic_memory to a longTermMemory config (OCG url, server- resolvable credential NAME, agent/user/scope, maxResults, summary model) plus a feedbackSink worker ref when a feedback_sink callable is set. - OCGMemoryStore gains a `credential` name (default OCG_PUBLIC_KEY) for the compiled path, distinct from the raw client `token`. - Register a best-effort feedback_sink SIMPLE worker that rebuilds a FeedbackEvent and delivers the good/bad capability links out-of-band. Server compiler: - New LongTermMemoryConfig DTO + AgentConfig.longTermMemory/feedbackSink. - compileWithTools inlines pre-loop retrieval (search -> format -> set _ltm_context) and post-loop distill -> build value -> save -> feedback links -> feedback_sink. All tasks optional so memory never fails the run. - Inject retrieved memory as an LLM system message (mirrors _human_feedback). - Reuse host-mode-aware credential header rewrite via ToolCompiler. - JavaScriptBuilder helpers to format search hits (folding good/bad signal) and parse the distiller JSON into the durable value string. - Default inbound-webhook `media` to [] to avoid an upstream NPE. Tests: config serializer coverage for longTermMemory + feedbackSink, absent no-op, and summary-model fallback. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
NicholasDCole
force-pushed
the
feature/add_memory
branch
from
July 8, 2026 01:25
d61ee46 to
92cad96
Compare
GET /agent/executions/search accepts topLevelOnly (default false). When true, the service appends `parentWorkflowId = ""` to the query (AND-joined when a query is already present) so only root executions are returned, excluding sub-agent/sub-workflow rows. Backed by the queryable parent_workflow_id column in orkes-conductor. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
# Conflicts: # server/conductor-agentspan/src/main/java/dev/agentspan/runtime/compiler/ToolCompiler.java # server/conductor-agentspan/src/main/java/dev/agentspan/runtime/controller/AgentController.java # server/conductor-agentspan/src/main/java/dev/agentspan/runtime/service/AgentService.java
…er + topLevelOnly) The main merge left the controller calling a boolean `topLevelOnly` variant while the service only had main's `String classifier` variant — a compile mismatch that also risked dropping one feature. Support both: AgentService gains a combined overload `searchExecutionsRaw(..., String classifier, boolean topLevelOnly)` that folds the classifier filter (via the existing withClassifierFilter — unchanged, tests intact) and ANDs in `parentWorkflowId = ""` when topLevelOnly. The 5-arg and (…, classifier) overloads delegate to it. AgentController exposes both `classifier` and `topLevelOnly` query params on /executions/search. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This was referenced Jul 13, 2026
Open
NicholasDCole
added a commit
to conductor-oss/conductor
that referenced
this pull request
Jul 14, 2026
Ports the server side of agentspan-ai/agentspan#298: agents configured with longTermMemory get compiler-inlined memory steps — pre-loop retrieval (search HTTP -> format INLINE -> _ltm_context variable injected as an LLM system message) and post-loop distill/save/ feedback-links tasks, plus an optional feedback_sink SIMPLE worker task. All memory tasks are optional=true so a memory outage never fails the agent. Adaptations for this repo: credential headers resolve via ${workflow.secrets.NAME} (this host has no CredentialAwareHttpTask, so the inert #{NAME} form would never resolve), no __agentspan_ctx__ forwarding, and the distill model falls back to the agent's own model when summaryModel is unset (ModelParser.parse(null) throws). Also ports the topLevelOnly executions-search filter and the media input-template default that fixes the inbound-webhook NPE. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
…an feedback links
Add OCGMemoryStore (a synchronous MemoryStore HTTP adapter over the OCG BFF) and wire long-term memory into the agent runtime:
Adds Agent params semantic_memory/memory_summary_model/feedback_sink, the OCGMemoryStore.feedback_links() mint call, example 118, and unit tests.