← All work
MEMORY / RETRIEVAL2025–2026

Agent
Memory

Relevant memories, with time and contradictions accounted for.

PRODUCT DESIGN / EVALUATION

Case study by

FACTSTIMECONTEXTMEMORY
Memory design · facts, time and relationships

The problem.
My part in it.

Background

  • Sparse memory lost context; dense storage added duplicates, conflicts and retrieval noise.
  • Connections across time and validity checks for changed information.

Objective

  • Build the memory context needed for a query.
  • Account for time and conflicting information.

My role

  • Reviewed product needs, evaluation and design changes.
  • Compared retrieval approaches and future application.

How it connects.

Memory architecture
Memory writing and question-time retrievalStore facts with time and relationships; retrieve evidence for the current questionMemory writing and question-time retrievalStore facts with time and relationships; retrieve evidence for the current questionMEMORY WRITE PATHQUESTION-TIME RETRIEVALLATER DESIGN CHANGEStored memoryConversation — Facts and contextConversationFacts and contextExtract memories — Separate important informationExtract memoriesSeparate importantinformationReconcile facts — Changes and valid timeReconcile factsChanges and valid timeMemory store — Facts + relationshipsMemory storeFacts + relationshipsCurrent question — Question and conversation contextCurrent questionQuestion and conversationcontextRetrieve & assemble evidence — Relevant facts with contextRetrieve & assembleevidenceRelevant facts with contextResponse context — Evidence for generationResponse contextEvidence for generationReview entity-retrieval latency → atomic facts (TCR) and LLM links
Processing / data flowIteration / feedbackService / environment boundary

The outcome.

89.2%Answerable Success
446 / 500Questions answered successfully

Internal evaluation when temporal information was applied. Atomic facts and LLM links belong to a later design change.

NEXT PROJECT / 08Creative Studio