ODC Agent Workbench provides two built-in memory mechanisms — RAG-based grounding data and turn-by-turn conversation storage in an ODC Entity — that work well for short single-session interactions but break down for long-running multi-session agents. The limitations include growing context window pressure, no knowledge consolidation, no cross-session user memory, single-strategy retrieval, and opaque memory. Hindsight, an open-source memory system by Vectorize, addresses these gaps by extracting structured memories at ingestion time (World Facts, Experience Facts, Observations, Mental Models) and retrieving them via four parallel strategies (semantic, keyword, graph, temporal) fused through its TEMPR system. The article explains how to integrate Hindsight into an OutSystems agent flow using Recall before LLM calls, Retain after responses, and tag-based scoping (user:, agent:, session:) for multi-user safety. Trade-offs including added latency, cost, architectural dependency, and reduced transparency are also covered, along with a walkthrough of the Hindsight Demo Agent available on OutSystems Forge.
Table of contents
How Memory Works Today in ODC Agent WorkbenchGrounding DataConversation MemoryWhere the Default Falls ShortShort, Single-Session InteractionsLong-Running, Multi-Session Agents with User ContextWhat Hindsight Is — and Why It FitsHindsight Memory Concepts Deep DiveWorld FactsExperience FactsObservations and Mental ModelsHow It All Fits TogetherAt the start of every LLM call: World facts and experience facts go into the promptDuring the conversation: Observations and Mental Models via ReflectAfter the LLM responds: New memories are retainedScoping Memory with TagsThe Three Tag Scopes Worth KnowingGet Stefan Weber’s stories in your inboxCombining TagsTrade-Offs and ConsiderationsAdded Architectural DependencyLatencyCostDebugging and TransparencyNot a Full Replacement for All ScenariosAttaching a Hindsight Memory Bank to an OutSystems AgentPrerequisitesCreate a Memory BankAdd Documents to the Memory BankConfigure Agent in ODC PortalHaiku ModelImplementationLoadKnowledge ActionLoadUserExperience ActionBuildMessages ActionStoreMemory ActionTry the Demo AgentSummarySort: