Document Knowledge
Retrieved source material from uploaded documents. Good for policies, runbooks, and reference content.
Memory is how ThinkWork carries useful context forward without treating every backend as the product.
The clean model is:
That means Memory is not just one store, and it is not just long-term recall. It is the harness-owned context layer above changing retrieval and memory engines.
Memory in ThinkWork spans four related concerns:
Document Knowledge
Retrieved source material from uploaded documents. Good for policies, runbooks, and reference content.
Long-term Memory
Carry-forward context learned from prior work. ThinkWork exposes one memory contract above pluggable engines.
Retrieval and Context
How thread history, retrieved documents, and recalled memories come together inside a turn.
Structured Memory Direction
Forward-looking direction for richer entity and relationship-aware memory. Roadmap, not current core product surface.
Today, the stable story should be read in this order:
ThinkWork currently supports two long-term memory adapters:
Exactly one long-term memory engine is canonical per deployment.
Short-term and session context still come from thread history. Long-term memory should not masquerade as the thread record.
This keeps the product honest.
It avoids teaching people that:
The contract belongs to ThinkWork. Engines implement it.