One search. Connected ideas.

Long-term memory stores selected information beyond one task. Semantic memory contains facts or preferences; episodic memory records past events; user-profile memory contains explicit attributes such as preferred language. These categories overlap and should…

Beginner — Context versus memory?

Summarization compresses earlier interaction into a shorter record. Memory retrieval selects relevant stored entries, sometimes through vectors and metadata. Both reduce context volume but can lose or distort information.

Tools are approved external capabilities. Planning breaks a goal into steps; state records current execution facts; memory carries selected information across turns or tasks. Autonomous action means the system can perform some actions without a new user…

A checkpoint records execution progress and state at a recovery boundary. Persistence saves it outside the process. Memory is selected information reused by later turns or tasks; a checkpoint is not automatically good long-term user memory.

Context is what the model sees now. Short-term memory is task/session information retained across nearby steps. Conversation memory is selected dialogue history or a summary used to maintain continuity.

MongoDB collections can store conversation messages, task/run state, approved preferences, and memory records. A LangGraph checkpointer requires its expected integration contract rather than an arbitrary collection called memory. Business claims remain…

- Deterministic: known rules and transitions. Best for permission, arithmetic, approvals, and business commits.

vLLM is an inference-serving system designed for efficient model execution. Serving techniques such as batching and attention-memory management help multiple requests share expensive hardware. GPU optimization is matching workload, memory, scheduling, and…

State persistence stores task progress; checkpoints record recovery positions. Neither automatically means user preference memory. A durable claim workflow may need execution records even when conversational history is deleted under a different retention…

Quantization represents model values with fewer bits to reduce memory and sometimes improve speed. GGUF is a model file format used by compatible local inference runtimes; it is not itself a model or a guarantee of quantization level. Local embeddings run an…

PEFT means parameter-efficient fine-tuning: update a smaller subset of parameters instead of the whole model. LoRA trains small low-rank adapter matrices attached to selected model operations. QLoRA combines adapter training with a quantized frozen base model…

Long-term memory stores selected information beyond one task. Semantic memory contains facts or preferences; episodic memory records past events; user-profile memory contains explicit attributes such as preferred language. These categories overlap and should…

Summarization compresses earlier interaction into a shorter record. Memory retrieval selects relevant stored entries, sometimes through vectors and metadata. Both reduce context volume but can lose or distort information.

Beginner — Context versus memory?

Tools are approved external capabilities. Planning breaks a goal into steps; state records current execution facts; memory carries selected information across turns or tasks. Autonomous action means the system can perform some actions without a new user…

Context is what the model sees now. Short-term memory is task/session information retained across nearby steps. Conversation memory is selected dialogue history or a summary used to maintain continuity.

vLLM is an inference-serving system designed for efficient model execution. Serving techniques such as batching and attention-memory management help multiple requests share expensive hardware. GPU optimization is matching workload, memory, scheduling, and…

A checkpoint records execution progress and state at a recovery boundary. Persistence saves it outside the process. Memory is selected information reused by later turns or tasks; a checkpoint is not automatically good long-term user memory.

MongoDB collections can store conversation messages, task/run state, approved preferences, and memory records. A LangGraph checkpointer requires its expected integration contract rather than an arbitrary collection called memory. Business claims remain…

I define which sources may create memory, require provenance and verification for important facts, scope it by tenant/user, and provide expiry and correction/deletion. Retrieved memory remains lower authority than fresh canonical business data.

State persistence stores task progress; checkpoints record recovery positions. Neither automatically means user preference memory. A durable claim workflow may need execution records even when conversational history is deleted under a different retention…

It must resolve current policy applicability through the business service. Historical memory may explain past conversation but cannot decide the active version. I invalidate dependent caches and test stale-memory conflicts explicitly.

Context is the information a model sees in the current call. Memory is stored application data that may be selected for later context. Persisting something does not make it accurate, current, relevant, or authorized for every task.

AI Engineering learning module

AI Engineering learning module

AI Engineering learning module

AI Engineering learning module

AI Engineering learning module

AI Engineering learning module

AI Engineering learning module

AI Engineering learning module