Why Does AI Always Misremember? Understanding LLM Memory Hallucination
Context Window Is Not Memory LLMs compress inputs into vectors and recall knowledge via probability sampling, not database queries. When context grows long o…
Context Window Is Not Memory
LLMs compress inputs into vectors and recall knowledge via probability sampling, not database queries. When context grows long or information appears early, models forget because attention mechanism assigns lower weight.
Lost in the Middle Effect
A 2023 paper showed models perform best with info at the beginning or end, worst in the middle. We hit this at SFD Lab: constraints in paragraph 3 of a prompt were ignored; moving them to paragraph 1 fixed the problem immediately.
Cross-Session Memory Does Not Exist
Most models have no persistent memory across conversations. Products claiming AI remembers preferences use: vector database recall, summary compression, structured storage, or system prompt pre-filling. Our OpenClaw reads MEMORY.md and daily logs before each conversation, which is engineered pseudo-memory, not real remembering.
Why Models Fabricate Memories
RLHF training pushes models toward satisfying answers. When users ask about a previous discussion, the implicit expectation is a positive answer, and models learned to accommodate. This is a real danger in medical, legal, and financial contexts.
Practical Strategies
- Put critical info first, never bury it in the middle
- Do not ask do you remember, just repeat the requirement
- Break long tasks into steps with focused context
- Use files for persistence, not conversation history
- Have the model restate key requirements before starting
SFD Lab Note: Our OpenClaw Agent uses the file system for memory. Daily logs go to memory/YYYY-MM-DD.md, decisions to MEMORY.md. Not depending on model context memory is foundational to our stable team operations.