Knowledge, context
and continuity.

I treat retrieval as a way to select evidence for generation. RAG brings retrieved information into an answer; storage keeps the material available. Direct access to files, search and structured queries are architecture choices I can combine around the task and the evidence required.

In my RAG experiments, relevant chunks existed in the index but were absent from the retrieved results: a retrieval failure. I examine the query and retrieval path before treating an incomplete answer as missing knowledge. Source references help me check what actually reached the model.

The working context is the information available to the model for the current step. I want it to include the task, the relevant sources and the decisions that still matter. Accumulating more material does not automatically make that context more useful.

Assessing many risks and rationales together proved less reliable in my experiments. I explored individual evaluations without carrying earlier judgments into the next assessment, using prompt chains to run the steps in sequence. Context rot and task complexity are possible explanations, but that experience does not isolate a cause.

Markdown can hold readable notes; JSON can describe structured state. PostgreSQL can store documents, metadata and JSON alongside other records. Those formats and storage choices can work together.

Persisted task state has a separate job: carrying decisions, completed steps and unresolved questions beyond the current context window. Thread checkpoints and longer-term stores can support different forms of continuity. I want a resumed task to recover the work and its sources, with enough detail to see what still needs review.