Skip to main content
Need help?
Ask in DiscordOpen a GitHub issueSupport guide

​
Summary

This starter shows one clean way to separate active notes, working memory, imported personal context, retrieval inputs, citation-bearing verifiable RAG, and durable artifacts in a small agent loop.

​
Status

starter Source code: patterns/examples/agent-memory-retrieval-starter

​
Why It Exists

Memory examples often collapse everything into one vague history object. This starter is intentionally narrower: it highlights the boundary between what the agent is currently holding, what a user imported from another assistant, what it can retrieve, what must remain an external file-search corpus, and what it should preserve as an artifact. The May 2026 multimodal file-search updates from Google made that retrieval boundary easier to teach. A useful starter should show metadata filters, page-level citations, and multimodal chunks without pretending that retrieval stores are the same thing as agent memory.

​
Folder Structure

agent-memory-retrieval-starter/
├── index.mdx
└── src/
    ├── artifact_policy.py
    ├── memory_flow.py
    ├── personal_context.py
    ├── retrieval_trace.py
    └── verifiable_rag.py

​
Quick Start

From the repository root: 
python3 scripts/verify_example_projects.py
This is still a starter, not a full application. Read src/memory_flow.py for the state boundary and src/verifiable_rag.py for the retrieval-side citation surface.

​
Included Sample Files

  • src/memory_flow.py: the smallest useful state container for active notes, task-scoped working memory, imported personal context, retrieval inputs, and durable artifacts
  • src/personal_context.py: a tiny normalization surface that keeps imported preference facts separate from retrieval results and artifacts
  • src/retrieval_trace.py: a tiny ranking and trace surface for making retrieval decisions inspectable
  • src/verifiable_rag.py: a tiny filter-and-citation surface for metadata filters, page numbers, and multimodal grounding output
  • src/artifact_policy.py: one place to show how a starter can separate artifact-promotion rules from raw note capture

​
Constraints

  • No storage backend is wired yet.
  • No embedding or vector-store integration is included.
  • No provider SDK is called directly.
  • Imported personal context stays deliberately small and reviewable.
  • The example favors clear boundaries over completeness.
  • Citations are modeled as output artifacts, not as proof that retrieval is always correct.

​
Flow Boundaries

The starter may:
  • keep working memory and personal context separate from retrieval inputs
  • apply metadata filters before selecting chunks
  • surface page-level and media citations as part of the returned artifact
  • keep retrieval traces inspectable enough for review
The starter must not:
  • treat a file-search store as durable agent memory
  • silently promote retrieved snippets into long-term memory
  • hide filter decisions or cited locations from the reviewer

​
Next Steps

  • Add a simple persistence layer for imported context and durable artifacts.
  • Add a clearer review workflow for memory imports before they become active.
  • Add a provider adapter for one real file-search API without collapsing the retrieval store into the memory layer.