While meeting up with a friend, we talked about the kinds of things we’ve built using AI or tools to assist us. They mentioned they’d built their own knowledge base MCP server for Claude Code — no code, no write-up, just “I made a thing that gives Claude memory.” That was enough. Claude Code sessions are stateless, and I’d been annoyed by it for a while.
Every new session started from zero. I’d re-explain the same architecture, re-discover the same gotcha, watch Claude confidently re-propose an approach we’d already ruled out the day before, and it always got dropped by the next /clear. So I sat down and built Knowledgebase, my own take on the idea: a persistent, per-project memory for Claude Code.
What it actually does
The whole thing is a local SQLite database (~/.claude/knowledgebase.db) with full-text search, wired into Claude Code’s hook system so capture and recall both happen automatically instead of depending on anyone — me or Claude — remembering to take notes:
UserPromptSubmitruns on every prompt. It searches the current project’s saved entries and injects a<kb-context>block into the conversation before Claude even starts working. This is the recall half — it happens whether or not anyone asks for it.- MCP tools (
kb_add_entry,kb_search,kb_get_project, and a few others) let Claude read and write knowledge explicitly, mid-session, when something worth keeping comes up. Stopfires at the end of every session as a backstop, nudging a last capture pass before the session closes.PreCompact/SessionEndauto-save a summary on/compactand/clear— the two moments where context disappears before Claude gets a chance to save anything itself.
The Stop hook is the one I almost skipped — it felt redundant next to the MCP tools. But in practice Claude doesn’t always remember to write things down mid-session, so having something nag it on the way out catches more than I expected.
The decision I spent the most time on
The auto-save feature on /clear//compact needs something to actually write the summary, and the obvious choice was to shell out to claude -p and let the CLI do the summarizing. I went back and forth on this one more than I expected to.
Knowledgebase is meant to work as a general tool, not something permanently welded to whichever chat client I happen to be using today. So the summarizer is pluggable: it looks for KB_SUMMARIZER_CMD first — any shell command that reads a transcript on stdin and writes back {"title", "body", "tags", "type"} on stdout — and falls back to calling the Anthropic API directly over HTTPS (not the CLI) if ANTHROPIC_API_KEY is set. If neither is configured, it just no-ops, quietly, and you’d never notice unless you went looking for the summaries that never showed up.
It also tracks its own progress — a small hook_state table keyed on transcript path and last line read — so a long session that triggers /compact five times doesn’t produce five overlapping summaries. Each firing only summarizes what’s new since the last one.
Running your own tooling has its own bugs
At one point I looked into why /clear wasn’t triggering an auto-save at all, and the answer wasn’t in the summarizer logic — it was that the SessionEnd and PreCompact hooks were never actually registered. My dotfiles repo pins this project as a submodule, and that submodule was still checked out at a commit from before those two hooks existed. The install script was working perfectly; it was installing a version of the project that was already out of date.
It’s a small thing, but it was a good reminder that building your own dev tooling doesn’t exempt you from normal dev tooling problems — stale pins, silent no-ops, config drift. The fix was boring: bump the submodule, rerun install.sh.
Try it yourself
If you want the same thing:
git clone https://github.com/jsundquist/knowledgebase
cd knowledgebase
./install.shPrerequisites are just Python 3.11+ and uv. The install script runs uv sync, registers the MCP server in ~/.claude.json, and wires all four hooks into ~/.claude/settings.json. It’s safe to re-run if you update later. Restart Claude Code afterward, and it’s live — no further setup needed for the core recall/capture loop.
The auto-save-on-/clear feature is opt-in on top of that. If you want it, set either KB_SUMMARIZER_CMD to point at your own summarization script, or set ANTHROPIC_API_KEY to let it call the Anthropic API directly with a small model. The second option means a transcript excerpt leaves your machine every time you /clear or /compact, which is why I made the whole feature do nothing until you’ve explicitly picked one.
What it’s gotten me
Sessions pick up where the last one left off without me doing anything special — a decision made on Tuesday is still known on Thursday, a dead end I already ruled out doesn’t get re-proposed, and the parts of a project that used to be tedious to re-explain just aren’t anymore.
All from one sentence a friend mentioned in passing about a thing they’d built for themselves.