A Self-Updating Knowledge Base for Terminal AI
I spend most of my day in the terminal with an AI coding assistant.
I often solve hard problems during a session. I find a tricky fix or a specific config setting. Then I close the tab and the knowledge vanishes. A month later, I solve the exact same problem again.
I built a system to make my assistant maintain its own notes using Claude Code hooks.
The system uses three parts:
- Search a small Markdown knowledge base on every prompt to provide context.
- Capture useful notes when a session ends.
- Load the index when a session starts.
Here is how to build it without slowing down your workflow.
- Use UserPromptSubmit for retrieval
This hook runs every time you send a prompt. It receives your text and can inject context before the model responds.
Do not let the model decide when to look at your notes. Use a fast search like grep to find relevant files. Inject only the top five matches. This keeps the process under 100ms.
Two rules for search hooks:
- Keep it cheap. Use grep instead of an LLM for the initial search.
- Keep it lean. Only inject titles and file paths. The model can open the file if it needs more detail.
- Avoid the Stop hook for heavy tasks
The Stop hook runs every time the assistant finishes a reply. If you run a long task here, your assistant becomes slow. A 30-turn session will trigger 30 heavy tasks.
Instead, use SessionEnd to capture the full session once.
- Solve the SessionEnd limitation
SessionEnd is non-blocking. If you start a long background task here, the system might kill it when the session closes.
The fix is to use two hooks:
- SessionEnd: Quickly add the transcript path to a queue file.
- SessionStart: Drain that queue and run the heavy capture task in the background.
Because the next session is active, the background process will survive.
- Prevent infinite loops
When you spawn a headless AI to write notes, it inherits your environment. This means the new AI will trigger its own hooks. This creates a loop of endless AI calls.
Always add a recursion guard at the top of your scripts:
[ -n "$KB_CAPTURE" ] && exit 0
Set KB_CAPTURE=1 when you launch your background capture task.
Summary of the setup:
- Use UserPromptSubmit for instant context injection.
- Use SessionEnd to enqueue tasks.
- Use SessionStart to run the actual work.
- Use a cheap model for summarizing notes to save money.
One avoided re-solve pays for many captures.
Optional learning community: https://t.me/GyaanSetuAi