Thread by @ChrisLaubAI

Everyone talks about AI “memory,” but nobody defines it.

This paper finally does.

It categorizes LLM memory the same way we do for humans:

• Sensory
• Working
• Long-term

Then shows how each part works in GPTs, agents, and tools.

Here's everything you need to know:

First, what the survey does:

• Maps human memory concepts → AI memory
• Proposes a unifying 3D–8Q taxonomy
• Catalogues methods in each category
• Surfaces open problems + future directions

Think of it as a blueprint for how agents can remember.

Human ↔ AI memory parallels (Figure 1)

• Sensory → perception layer: raw input briefly held, then dropped or passed on.
• Working memory: immediate reasoning & dialogue context.
• Long-term memory: split into explicit (episodic + semantic) and implicit.

Explicit memory:

Episodic → Non-parametric, long-term (e.g. your preferences in a DB).
Semantic → Parametric, long-term (facts stored in model weights).

Implicit memory: learned skills/patterns, partly parametric, partly non-parametric.

The organizing frame: 3D–8Q taxonomy (Table 1).

Memory is classified along 3 axes:

• Object: Personal vs System
• Form: Non-parametric vs Parametric
• Time: Short- vs Long-term

8 quadrants, each with a role.

Quick rundown of the quadrants:

QI: Personal, non-parametric, short-term → dialogue context.
QII: Personal, non-parametric, long-term → user-specific history.
QIII: Personal, parametric, short-term → prompt/context caching.
QIV: Personal, parametric, long-term → knowledge editing.

Personal memory (Sections 3.1–3.2): how the system remembers you.

It captures your inputs and preferences → boosts personalization.
Two forms:

Non-parametric (external DB, graphs, vectors)

Parametric (internalized inside the model)

QI: Non-parametric, short-term.

Classic chat memory: using recent turns for coherent, intent-aware responses.

Example: context window in GPT-style chatbots.

QII: Non-parametric, long-term.

Cross-session memory: storing user history for personalization.

Pipeline has 4 steps:

• Construction
• Management
• Retrieval
• Usage

Benchmarks exist for eval.

QIII: Parametric, short-term.

Caching contextualized chunks or attention states.

Helps with RAG + multi-turn settings.

Lower cost, faster response.

QIV: Parametric, long-term.

Personal traits embedded into model parameters.

Methods: fine-tuning, editing.

Bucket = knowledge editing.

System memory (Section 4): how the agent remembers its own work.

Definition: intermediate states generated during problem-solving.
Key use: improve reasoning, planning, and self-evolution.

QV: Non-parametric, short-term.

Reasoning + planning traces.
Think: ReAct, Reflexion.

QVI: Non-parametric, long-term.

Reflections consolidated into reusable workflows.

Examples: BoT, AWM, TiM, Voyager, Retroformer, ExpeL.

QVII: Parametric, short-term.

Efficiency layer.

KV-cache management, compression, quantization.
Systems: vLLM, ChunkKV.

QVIII: Parametric, long-term.

Knowledge integrated into weights over time.

Frameworks: MemoryLLM, WISE.
Designs: dual-memory routing, self-updating LLMs.

Why this taxonomy matters

It gives builders a checklist:

Whose memory is this for?
Where does it live?
How long should it last?

That clarity = intentional design.

Open problems (future directions):

• Multimodal memory (text + image + audio + video)
• Stream memory (real-time, continuous)
• Comprehensive integration of all memory types
• Shared memory across models
• Group-level privacy
• Automated self-evolution

Practical takeaways:

• Design against the 3 axes: object, form, time.
• Always pair QV (trace) with QVI (reflection).
• Engineer for retrieval first.
• Budget for speed (KV caches).
• Be deliberate with editing.

This paper = the clearest roadmap to AI memory so far.

Read the full paper: arxiv.org/abs/2506.02153…

P.S.

Are you running a business?

AI can help you automate all your tasks for you.

Join us here at Sentient Automations and learn how to scale with AI for free:
whop.com/sentient-ai-au…

I hope you've found this thread helpful.

Follow me @ChrisLaubAI for more.

Like/Repost the quote below if you can:

https://twitter.com/1598365749940310019/status/1960687486532976908