IIT Madras article splits LLM memory into five approaches and asks why belief revision is still missing
LimeOptimal539 · reddit · 2026-10-20
A Hugging Face community article by R. Vimalkanth and T. Chudiwal (IIT Madras) proposes a five-way framework for LLM memory, centered on why long context and retrieval aren't enough for belief revision when new information conflicts with old state.
The five categories:
- Store more: keep explicit history (KV caches, RAG); old and new coexist, resolved at read time at growing cost.
- Compress: fold history into a fixed-size state (linear attention, RetNet, Mamba); saves space but stored associations interfere.
- Correct: gate or prediction-error updates control what changes (gated linear attention, mLSTM, DeltaNet); more efficient but hard to scale to long contexts.
- Learn at inference: update mutable parameters from incoming evidence (TTT, Titans).
- Distribute plasticity: spread changes across sparse, high-dimensional interacting state (Dragon Hatchling / BDH).
The author illustrates with a coding agent learning a function signature changed: keeping both definitions retains the info but still requires read-time arbitration. They also question whether all five categories belong at the same level of comparison, since some are model-level mechanisms and others system-level designs.
More from coding & agent
- OpenAI's dots can now use company tools via Merge for Workforce with employee-bound permissions — shensi · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Agents with full-disk and accessibility access are an enterprise security nightmare, argues Runlayer — leveragedupside · 2026-10-21
- Acellera runs an unsupervised virtual lab of AI scientists on a real IL-23 molecular discovery program — gdefabritiis · 2026-10-21
- Pydantic author experiments with JS/TS and React support in the Monty sandbox — samuelcolvin · 2026-10-21
- Supabase lets coding agents query your logs to catch errors and slow queries — dshukertjr · 2026-10-21