Memory Types · Taxonomy
Parametric vs Non-Parametric Memory in AI
Parametric memory is knowledge stored in model weights (fine-tuning); non-parametric memory is external — vectors, graphs, databases — that agents read and write at runtime.
Parametric
In model weights · fine-tuning · opaque
Non-parametric
External store · vectors/graphs · editable
Parametric
What is parametric memory?
Knowledge encoded in neural network weights — updated via training or fine-tuning, not at runtime.
In-weight memory: fast inference (no retrieval step), but hard to update per-user, opaque and expensive to change. Domain fine-tuning encodes stable knowledge into parameters — medical terminology, company tone, product catalog patterns.
Non-parametric
What is non-parametric memory?
External stores — vector DBs, knowledge graphs, Redis, memory APIs — updatable without retraining.
This site’s focus. Engram, Mem0, Zep, Letta and LangMem all implement non-parametric memory: extract facts, store externally, retrieve at query time. Mem0 LOCOMO J 66.9; Zep 66.0 (Chhikara/Rasmussen et al., 2025).
Comparison
Parametric vs non-parametric: comparison table
| Dimension | Parametric | Non-parametric | Hybrid |
|---|---|---|---|
| Update mechanism | Retrain / fine-tune | Runtime read/write API | Both |
| Per-user facts | Not practical | Native (user_id scoping) | Non-parametric for users |
| Interpretability | Opaque weights | Inspectable facts | Mixed |
| Cost to change | High (GPU training) | Low (API write) | Domain in weights, facts external |
| Frameworks | Fine-tuning pipelines | Engram, Mem0, Zep, Letta | Production assistants |
| Best for | Stable domain knowledge | Dynamic user facts | Stateful production agents |
Frontier
Latent and token-level memory
Beyond text stores — representations inside the model itself (latent memory, token-level caching).
Agent-memory surveys (arXiv taxonomy) distinguish token-level, parametric and latent memory from external non-parametric stores. Research frontier — brief pointer only. See advanced cluster for continual learning and latent memory.
Decision
When to use which
- Parametric — stable domain vocabulary, tone, reasoning patterns that rarely change
- Non-parametric — per-user preferences, conversation history, evolving CRM facts
- Hybrid — fine-tuned base model + Engram/Mem0/Zep memory layer (most production assistants)
FAQ
Frequently asked questions
What is parametric memory in AI?
Knowledge encoded in model weights — updated via training or fine-tuning, not at runtime. Fast inference but opaque and hard to update per-user.
Is RAG parametric memory?
No — RAG is non-parametric. Documents are stored externally in a vector index and retrieved at query time, not baked into model weights.
Is Mem0 parametric memory?
No — Mem0 is non-parametric external memory (vector store + extraction API). Facts are written and retrieved at runtime without retraining.
Is fine-tuning parametric memory?
Yes — fine-tuning encodes knowledge into model weights. Compare with external memory in memory vs fine-tuning.
What is the hybrid parametric + non-parametric pattern?
Fine-tuned base model for domain knowledge + Engram/Mem0/Zep for per-user facts. Most production stateful assistants use this hybrid.
What is latent memory?
Representations stored inside the model (beyond external text stores) — a research frontier. See latent and token-level memory.
How does parametric memory relate to continual learning?
Continual learning updates weights over time (parametric). External memory avoids catastrophic forgetting by keeping facts outside weights. See continual learning vs memory.