External memory stores knowledge outside a model’s learned parameters in a structured, searchable corpus. A retriever gives the model access to that information on demand, like an open-book exam. section 6.1
Moving knowledge outside model parameters changes how it can be maintained and checked:
- Add, delete, or modify indexed documents without retraining the model; this can keep information current and make domain specialization cheaper. section 6.1
- Responses can cite source documents, but retrieval adds inference time and can fail to find relevant knowledge. section 6.1
Changing information
A financial service needs current market reports; a legal tool needs recent rulings; a clinical support system needs the latest medical research. section 6.1
Table 6.1 Parametric vs non-parametric knowledge
|
Feature
|
Parametric Knowledge (e.g., GPT-3)
|
Non-Parametric Knowledge (e.g., Atlas)
|
|
Knowledge Freshness
|
Static; frozen at the time of training.
|
Dynamic; updatable in near real-time.
|
|
Verifiability
|
Opaque; "black box" reasoning.
|
Transparent; responses grounded in citable sources.
|
|
Cost to Update
|
Extremely high (full model retraining).
|
Low (re-index external corpus).
|
|
Domain Specialization
|
Requires expensive fine-tuning or prompting.
|
Achieved by curating the external corpus.
|
|
Hallucination Risk
|
High when knowledge is outdated or missing.
|
Low; the model can state that knowledge is not found.
|
|
Failure Mode
|
Factual hallucination when knowledge is missing/outdated.
|
Retrieval failure; can return "knowledge not
found."
|
Compare external memory with parametric knowledge across freshness, verifiability, update cost, specialization, and failure modes. section 6.1

