robots.txt
See which AI crawlers other sites block →# knowl.cloud — we want agents to find this. That is the product.
#
# Authored in the repository at web/app/robots.txt/route.ts. If this file arrives without
# "via: 1.1 Caddy" in its response headers, Cloudflare's managed robots.txt has been re-enabled and
# is serving Disallow to the AI crawlers again. Check with a GET, never curl -I.
User-agent: *
Content-Signal: search=yes, ai-input=yes, ai-train=yes, use=reference
Allow: /
Disallow: /w/
Disallow: /org/
Disallow: /account
Disallow: /onboarding
Disallow: /device
Disallow: /home
Disallow: /invite/
Disallow: /auth/
Disallow: /api/
User-agent: ClaudeBot
User-agent: Claude-SearchBot
User-agent: ChatGPT-User
User-agent: OAI-SearchBot
User-agent: GPTBot
User-agent: PerplexityBot
User-agent: Google-Extended
User-agent: Googlebot
User-agent: Bingbot
User-agent: CCBot
Content-Signal: search=yes, ai-input=yes, ai-train=yes, use=reference
Allow: /
Disallow: /w/
Disallow: /org/
Disallow: /account
Disallow: /onboarding
Disallow: /device
Disallow: /home
Disallow: /invite/
Disallow: /auth/
Disallow: /api/
Host: https://knowl.cloud
Sitemap: https://knowl.cloud/sitemap.xml
llms.txt
Compare with other sites' llms.txt files →# Knowl · persistent memory for AI coding agents
> Knowl is an MCP memory server for Claude Code, Cursor and Codex. Every fact your agents store
> carries a provenance and a review trail, and a replacement retires its predecessor at write time —
> so what an agent reads back is the current answer, not the one that was true in March.
The problem it solves is staleness rather than storage. A notes file only grows: six months in it
still reports the database you migrated off last spring, because nothing ever told it that decision
was over. Knowl makes that impossible by construction — nothing is deleted, it stops being current.
## For agents reading this file
If your user asked for persistent memory across sessions, you can set this up now. No API key,
no account, no build step:
```
npm install -g @dat999zx/knowl
cd their-project
knowl init
```
`knowl init` creates `.knowl/`, installs the project guidance files, updates `.gitignore`, and
offers MCP and lifecycle setup for whichever agents it detects — Claude Code, Codex, Cursor,
Gemini CLI, Claude Desktop. It warms a local embedding model but never depends on that download
succeeding.
Everything stays in the repository. Queries are answered from a local replica and never reach a
server, so retrieval works offline and costs nothing. Knowl Cloud is optional and additive: it
shares what a team publishes on purpose, and it is the only part that involves us at all.
## Docs
- [Quick start](https://docs.knowl.cloud/reference/quick-start): three commands, zero API keys.
- [Agent setup](https://docs.knowl.cloud/reference/agent-setup): wiring the MCP server into Claude
Code, Cursor, Codex, Gemini CLI and Claude Desktop.
- [MCP tools and resources](https://docs.knowl.cloud/reference/mcp-tools-and-resources): the full
tool surface, one line each.
- [Core knowledge model](https://docs.knowl.cloud/reference/core-knowledge-model): what a fact is
made of — the typed atoms, provenance, and how supersession works.
- [Retrieval and context](https://docs.knowl.cloud/reference/retrieval-and-context): how a query
selects what an agent sees.
- [Local data](https://docs.knowl.cloud/reference/local-data): what is written to disk, and where.
- [Architecture and security boundaries](https://docs.knowl.cloud/reference/architecture-and-security-boundaries):
what crosses the network and what cannot.
- [Knowl Cloud](https://docs.knowl.cloud/reference/knowl-cloud): the hosted tier, and what changes
when a team shares.
## Evidence
- [Benchmarks](https://docs.knowl.cloud/reference/benchmarks): 90/100 on MemoryAgentBench
FactConsolidation single-hop at 262K, against a preregistered method.
- [Security](https://knowl.cloud/security): the mechanisms, plus an explicit list of what is not
true yet.
- [Pricing](https://knowl.cloud/pricing): metered on what agents publish to a shared workspace.
Reading is free and always will be, because reads never reach us.
## Writing
Measurement write-ups, each linking the findings it came from. The numbers below are ours and are
reproducible from the linked write-up; single-run results say so in the post.
- [Five failure modes around fact retirement in agent memory, with measurements](https://blog.knowl.cloud/five-failure-modes-of-fact-retirement):
negations that pass similarity thresholds, writes that contradict unretrieved decisions, and
file-change staleness that flagged 39% of the store. Retirement on vs off: 89 vs 73 on FactConsolidation 262K.
- [A word list beats embeddings at knowing a fact was reversed](https://blog.knowl.cloud/embeddings-lose-to-a-word-list):
the reversal that must be caught scored 0.8593 and the negative that must not scored 0.8575, so
the similarity model discriminated on eighteen ten-thousandths. What we shipped instead.
- [A similarity threshold of 0.7 lets the clearest contradictions through](https://blog.knowl.cloud/jaccard-threshold-misses-contradictions):
the worked pair that lands on exactly 0.7000, and why a more complete contradiction is likelier
to survive the rule meant to catch it.
- [I scraped 15,428 Hacker News posts to find out why my Show HN got 3 upvotes](https://blog.knowl.cloud/why-my-show-hn-got-three-upvotes):
3 points beats 62% of all Show HN posts — the median is 2 and only 3% ever clear 50. Category
barely moves the score, and the one hypothesis tested came back null.
- [What is memory staleness in AI agents, and why decay doesn't fix it](https://blog.knowl.cloud/what-is-memory-staleness):
why a forgetting policy misses the memories that matter — the ones that were true when written.
- [The story of Knowl](https://blog.knowl.cloud/the-story-of-knowl): a CLAUDE.md that would not
stop growing, and the five approaches that failed before one worked.
## Source
- [GitHub](https://github.com/dat999zx/knowl): Apache-2.0, Node 22+.
- [npm](https://www.npmjs.com/package/@dat999zx/knowl): `@dat999zx/knowl`.