Document Knowledge

Document knowledge is the part of Memory that starts from uploaded source material rather than prior interactions.

Use it when the answer should come from documents your organization already owns, not from what an agent happened to learn during earlier work.

What it is good at

Document knowledge is a good fit for:

product docs and internal runbooks
policies and operating procedures
FAQ-style support content
long-form reference material that changes occasionally

This is a grounding layer, not a replacement for memory.

Document knowledge answers: “What does the organization know from source material?”
Long-term memory answers: “What has this agent or team learned across work?”

Where it fits in ThinkWork

The full model is:

Threads store what happened
Document knowledge provides retrieved source material
Long-term memory carries learned context forward
Retrieval and context assembly decide what reaches the turn

That is why document knowledge lives under Memory. It is one context source, not the whole story.

Current implementation

Today, ThinkWork’s document-backed retrieval path uses Bedrock Knowledge Bases.

The rough flow is:

Document upload
  -> S3 knowledge bucket
  -> Bedrock Knowledge Base sync
  -> embeddings + vector index
  -> retrieval during agent turns

The source files stay in S3 in your AWS account. Retrieval runs through Bedrock Knowledge Bases backed by Aurora Postgres with the pgvector extension — the same Aurora cluster that stores agents, threads, and memories. There is no separate vector database and no OpenSearch Serverless collection to pay for.

Operational notes

document sources do not automatically resync when files change
chunking strategy is set when the knowledge base is created
vector index lives in Aurora (pgvector), so there is no additional vector-store cost floor — you pay for the Aurora cluster you already have

What this page is not

This page is about document retrieval as a context source. It is not the full memory contract, and it is not the canonical record of work.

For that broader model, start with: