Long-term Memory
Long-term memory is how ThinkWork carries useful context forward from prior work without turning the thread record into the only recall mechanism.
Unlike document knowledge, which starts from uploaded source material, long-term memory starts from what the system learned across interactions.
What long-term memory means in ThinkWork
Section titled “What long-term memory means in ThinkWork”Long-term memory is for context like:
- user or team preferences
- ongoing project context
- prior decisions and follow-up items
- durable facts worth recalling across threads
- summaries or abstractions of earlier work
The important product rule is simple:
- Threads are the canonical record of work
- Long-term memory is the selective carry-forward layer
The ThinkWork memory contract
Section titled “The ThinkWork memory contract”ThinkWork owns the memory contract above any one backend.
That means ThinkWork should provide the stable semantics for:
- retain
- recall
- inspect
- export
- context assembly
The underlying engine can vary.
Today, ThinkWork supports two long-term memory adapters:
Exactly one engine is canonical per deployment.
- Hosted ThinkWork defaults to Hindsight
- Self-hosted and serverless-friendly deployments may choose AgentCore Memory
Short-term context versus long-term memory
Section titled “Short-term context versus long-term memory”This distinction matters a lot:
- Short-term and session context come from thread history
- Long-term memory provides selective recall across prior work
Long-term memory should help the system avoid rereading everything. It should not pretend to replace the thread record.
Current engine choices
Section titled “Current engine choices”Hindsight is the richer long-term memory path and the default for hosted ThinkWork.
It is the better fit when you want:
- stronger inspectability
- a clearer portable memory plane
- richer recall behavior
- future graph and reflection direction
AgentCore Memory is a valid long-term memory option for deployments that want a serverless-first path inside AWS-managed infrastructure.
It is a good fit when you want:
- AWS-native managed memory primitives
- lower infrastructure sprawl
- a simpler self-hosted/serverless deployment story
Capability differences may exist between engines. ThinkWork should surface those honestly instead of pretending every engine is identical.
Long-term memory versus document knowledge
Section titled “Long-term memory versus document knowledge”Use long-term memory for information learned from work. Use document knowledge for information loaded from source material.
Many agents use both:
- document knowledge grounds answers in docs and policies
- long-term memory carries forward learned context from prior work