한국어

Dawell Service Knowledge Management System
Purpose-built for knowledge management — Knowclaw

We built a self-growing organizational knowledge asset (an LLM wiki).
An autonomous agent that organizes scattered files & mail, grows them into a wiki, and answers from that wiki

Written 2026-06-01 · For partners & internal use · v1.0
KNOWCLAW
1 · In one sentence

An autonomous agent purpose-built for knowledge management

The Dawell Service knowledge management system is an autonomous agent that organizes and classifies scattered files & mail on its own,
grows a wiki incrementally from that content, and answers questions grounded in that wiki.

It is not a tool that digs through everything from scratch on every search. It settles what it reads into wiki pages, appends new material to the existing wiki as it arrives, and accumulates the wiki into a lasting organizational knowledge asset — closer to a knowledge librarian that gets smarter over time.

1 · Origin of the name

Why "Knowclaw"

One major thread in recent AI is the autonomous agent (agentic system). Give it a goal and it picks its own tools, decides, and asks back when stuck — the flagship being the general-purpose autonomous agent OpenClaw.

Knowclaw (Knowledge + Claw) takes OpenClaw's operating principle (a tool-use agent) as-is, but specializes it for one job: organizational knowledge management.

If OpenClaw is "the autonomous agent that does anything,"
Knowclaw is "the autonomous agent that grows an organization's knowledge best."
1 · What sets it apart

How it differs from a plain RAG tool

Most document Q&A re-searches the source on every question (RAG) — Knowclaw accumulates a wiki.

🗂️

It organizes on its own

Classifies scattered Drive files & mail by project and moves them into real folders.

→ Pipeline (Ch. 3)
📚

It grows a wiki

Reads the source to write person, concept, and summary pages, and accrues new material onto the existing wiki.

→ LLM wiki (Ch. 2)
📈

It gets smarter with use

As material piles up, the wiki thickens and answers grow more accurate.

→ Compounding effect (Ch. 6)
2 · Core idea

Not RAG re-search — a self-growing LLM wiki

Karpathy's "LLM Wiki" pattern — instead of re-searching every time, the LLM incrementally builds and maintains a wiki as a lasting asset.

RAG (retrieval-augmented)Knowclaw (LLM wiki)
Form of knowledgeRaw source-chunk embeddingsWiki pages refined & structured by the LLM
Handling a questionRe-searches the source each timeAnswers from the accumulated wiki
New materialJust adds chunks to the indexIncrementally integrated & linked into the wiki
Over timeIndex only grows (quality unchanged)The wiki thickens and quality rises
2 · Three-layer structure

Source → wiki → question/check

raw sourcesDrive files · mail (raw) → DriveFile (DB mirror + text cache)
↓  ingest (read & refine)
wikiLLM-written wiki pages → WikiPage (markdown + [[link]] graph)
↓  query / lint
query · lintWiki-grounded QA · quality checks

Wiki bodies stay in markdown, while storage, queries, and web rendering use Django DB models.

  • Better for multi-user access, web integration, and structured queries
  • Every wiki sentence's source is traceable (WikiPageSource)
3 · Pipeline

Four stages — collect · organize · wikify · ask

Whether run as a manual sync or a daily cron, it uses the same orchestrator (run_full_sync).

StageModuleWhat it does
📥 Collectdrive_syncRecursively scans project folders → diffs only changes → extracts text → upserts DriveFile
🗂 Organizeclassifier_agentdrive_organizerAgent designs the taxonomy & placement (plan) → a deterministic executor performs the actual Drive move
📚 Wikifywiki_agentReads organized sources closely to create/update/link wiki pages (integrated across all projects)
💬 Askwiki_qaNatural-language QA grounded in the wiki (multi-turn + source citations)
📧 +Mailemail_archiverLoads Gmail messages as files into project folders → reuses the same 4 stages
4 · Agent execution structure

Only reversible actions go to the agent — plan / execute split

We don't hand an irreversible action — moving a client's files — wholesale to on-the-fly LLM judgment.

✅ Reversible → agent tools

Taxonomy design & category-assignment plans (classifier), wiki page authoring & linking (wiki_agent). All happen on the DB, so they can be undone anytime.

⛔ Irreversible → deterministic executor

The actual move_file is performed by drive_organizer only through dry-run / apply / revert gates. The agent never holds this tool.

The agent does only reversible work (plans & text); irreversible Drive actions go to the deterministic executor.
4 · LangGraph tool-use standard

A shared in-house standard graph across three agents

Not a hand-rolled while loop — they share a common builder (build_tool_loop_graph) borrowed from the popax/MeetingClaw pattern.

AgentGraphCheckpointerTools (all reversible)
classifier_agentBatch, onceNonelist/read · propose/assign/merge category
wiki_agentBatch, onceNoneread source · read/write_wiki_page · link · search
wiki_qaMulti-turnPostgresSaversearch_wiki (embedding) · list · read_wiki_page

make_sequential_tool_node runs the tool_calls of one response in order, eliminating "one value per step" state conflicts at the root. Batch work needs no checkpointer; only multi-turn QA persists session context.

5 · Safeguards

Making an agent that touches client files trustworthy

① Reversible moves

Every move is logged to MoveLog → one revert restores it. Moves are confined below a designated root, deletion/trash calls are forbidden, and dry-run is the default.

② Project-boundary isolation

"Files never cross projects." An active guard ensures the move destination is always the current project folder.

③ Idempotency

Only top-level scattered files are moved (already-organized subfolders are preserved). A once-moved file is auto-excluded from the next sync.

④ Low-confidence mail held

If classification confidence is low or unmatched, it isn't loaded — only logged. The email stage is isolated with try/except to protect the core sync.

6 · Multiple projects

Isolated organizing, unified knowledge

🗂 Organizing is per-project, N times

It respects each client project's folder boundary. Even the same "contract" gets a separate category and folder per project.

📚 The wiki is one, across all

Its value is greatest when answering cross-project questions ("what issues do A & B share?"), so the wiki is a single unified knowledge base spanning all projects.

Physical folder organizing is isolated per project; knowledge search & QA are fully unified — merge the knowledge, don't mix the organizing.
6 · Impact

Quantitative impact — giving back the time spent organizing knowledge

Estimates that vary with volume of material and number of projects.

TaskBefore (manual)After (Knowclaw)Effect
Classifying Drive files & tidying foldersMinutes each × dozensAuto-classify & moveNearly automated
Writing summary/organizing docs30–60 min per itemAuto wiki buildMajor savings
"Where was that again?" searches5–15 minOne question + source citationMajor savings
Organizing & filing received mailMinutes eachAuto-loaded into project foldersNearly automated
6 · Key effect

A knowledge asset that gets better with use — the compounding effect

Pages ↑
More material → a thicker wiki
Links ↑
A denser knowledge graph
Accuracy ↑
Sharper answers to questions
Plain search finds it again every time;
a wiki grows with the organization.
7 · Tech stack · cost

Proven open source + multi-LLM, with cost visibility

Orchestration

LangGraph (StateGraph · tool loop · PostgresSaver) · LangChain · LangSmith

Reasoning & generation LLM

Anthropic Claude (classify · wiki · QA) · multi-provider swappable

Embeddings & search

OpenAI text-embedding-3-small (1536 dims) + numpy cosine

Web & data

Django 4.2 · DRF · gunicorn · PostgreSQL · psycopg 3

Shared & automation

standarda-core (Drive · Gmail · Docs) · crontab daily auto-sync

💰 Cost tracking

Tokens & USD auto-logged per run. Prompt caching cuts repeated input to 0.1× (measured −36% on Nulldam)

Design principle — "plan by agent, execute by deterministic executor" · "only reversible actions to the LLM"

What's next

Scaling embedding search

Once it grows to hundreds+ of pages, swap numpy cosine for pgvector.

Wiki .md export

Export the DB wiki to markdown files for external sharing & backup.

Stronger human triage

Let people review and re-classify low-confidence mail that was held from loading.

An organizational knowledge asset that gets smarter as material piles up

Thank you · chris@popupstudio.ai
navigate slides · F fullscreen · Home/End first/last