Context engineering.
Working memory.

Context engineering concerns the information available at each model call: instructions, tools, retrieved evidence and relevant history. Sourcegraph separates instructions, retrieval, memory and tools as a practical organizing model. The context therefore needs to be assembled for the current task. [01]

My working model separates stable instructions, authoritative live records, a small working-state note and knowledge retrieved on demand. I would read the current record through a tool, keep open questions and decisions in the working note, and load supporting documents when needed. This is my design synthesis rather than a required standard.

Anthropic describes compaction and structured note-taking for longer tasks. Compaction condenses a conversation; persistent notes can be read back into a later context. The tradeoff is what the summary retains: excessive compression can discard details that matter later. [02]

I want the saved state to include decisions, unresolved questions, the next action and links to actual outputs. A summary of an answer is not the saved result itself. I would keep source references with the result and reload the current state when work resumes.

Markdown and JSON are formats; PostgreSQL is a storage option; RAG is a way of retrieving evidence. They can work together. The choice depends on how information changes, how it is queried and what needs an authoritative record.

If a relevant chunk never reaches the model, I investigate retrieval. If the evidence is present but the answer misses it, I investigate context use and task design. If a completed result cannot be resumed because it was never saved, I investigate persistence. These need different checks.

Chroma’s Context Rot report found declining reliability as input length increased in its evaluated models and tasks. It supports testing how useful context is, with results tied to that evaluation. In my larger risk assessments, lower reliability remains an observation: context effects, conflicting material and task complexity are possible explanations to test. [03]