athena / knowledge
CLAUDE.md0 starsChanged 7 months agoMITView on GitHub
This is the second of Athena's two memories, and confusing it with the first is the single most expensive mistake you can make with a system like this. A recall system will never produce understanding, however much you put in it. That is not a defect, it is the design — MemPalace's own documentation says plainly that it does not summarise, extract, or paraphrase. Expecting compounding insight from it is like expecting a filing cabinet to write you a briefing.
In this repository57 other files it gives its agents.
- .claude/rules/agents.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- agent-harness-constructionSkill
- agentic-engineeringSkill
- ai-first-engineeringSkill
- api-designSkill
- backend-patternsSkill
- blueprintSkill
- bookmarkSkill
- brain-dumpSkill
- building-explainerSkill
- calendar-managerSkill
- code-engineerSkill
- compileSkill
- continuous-agent-loopSkill
- continuous-learningSkill
- copywriterSkill
- cost-aware-llm-pipelineSkill
- councilSkill
- database-migrationsSkill
- day-plannerSkill
- debriefSkill
- deep-researchSkill
- deployment-patternsSkill
- docker-patternsSkill
- e2e-testingSkill
- claude-apiSkill
- coding-standardsSkill
- eng-inboxSkill
- updateSkill
- enterprise-agent-opsSkill
- eval-harnessSkill
- frontend-patternsSkill
- humanizerSkill
- iterative-retrievalSkill
- log-expenseSkill
- memory-auditSkill
- migrateSkill
- minimalist-uiSkill
- onboardSkill
- full-output-enforcementSkill
- postgres-patternsSkill
- redesign-existing-projectsSkill
- regex-vs-llm-structured-textSkill
- researchSkill
- search-firstSkill
- security-reviewSkill
- security-scanSkill
- self-upgradeSkill
- high-end-visual-designSkill
- spending-summarySkill
- design-taste-frontendSkill
- tdd-workflowSkill
- transcribeSkill
- verification-loopSkill
- video-ideasSkill
- weekly-reviewSkill
The file
We split this CLAUDE.md into the rules it gives an agent
6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 43
Never edit a raw file afterwards.
- Very specific: line 46
Find the pages it touches: Check _index.md first. A source about pricing goes into the existing pricing page.
- Distinctive: line 49
Rewrite those pages: Not append. Rewrite. If the source refines a claim, the claim changes.
- Distinctive: line 64
When two sources genuinely disagree and you cannot tell who is right, say so in both pages:
- Distinctive: line 119
One page per subject: Always extend, never duplicate.
- Distinctive: line 121
Delete a page that duplicates a live file elsewhere: Do not lint it harder. A clean check on the wrong artefact is worse than a dirty one.
≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
The wiki — what you actually know
This is the second of Athena's two memories, and confusing it with the first is the single most expensive mistake you can make with a system like this.
| Recall | Synthesis | |
| Holds | what was said, verbatim, when, by whom | what it means |
| Answers | "what did Sarah say in June" | "what do I know about pricing" |
| Built by | storing and searching | reading and rewriting |
| Lives in | MemPalace | here |
A recall system will never produce understanding, however much you put in it. That is not a defect, it is the design — MemPalace's own documentation says plainly that it does not summarise, extract, or paraphrase. Expecting compounding insight from it is like expecting a filing cabinet to write you a briefing.
So sources land in raw/, and then somebody has to actually read them and rewrite what
changed. That rewriting is the whole point. A wiki that only accumulates is a folder.
Structure
knowledge/
raw/ immutable sources, never edited after capture
wiki/ topic pages, rewritten on every ingest
questions/ things you noticed you do not know
digests/ periodic roll-ups
_hot.md the ~500-token cache. Read this first, always
_index.md generated. Do not hand-edit
_log.md append-only record of every operationRead _hot.md before anything larger. It is a few hundred tokens and answers most
questions about where things stand. Crawling the whole wiki to answer "what am I working
on" is the expensive mistake this file exists to prevent.
The protocol, when a source comes in
- Save it raw.
raw/YYYY-MM-DD-<slug>.mdwith the URL, the date, and who made it. Never edit a raw file afterwards. It is the record of what was actually said, and you will want it when a page turns out to be wrong. - Find the pages it touches. Check
_index.mdfirst. A source about pricing goes into the existing pricing page. - Rewrite those pages. Not append. Rewrite. If the source refines a claim, the claim changes. If it contradicts one, both sides get stated (below). Appending is what turns a wiki into a pile of undigested quotes.
- Never create a second page for a subject that already has one. This is the failure mode. Two pages on the same thing means neither is trusted and both go stale. Extend the existing page.
- Park what you cannot resolve. A real disagreement between two credible sources is
a question, not a page.
questions/YYYY-MM-DD-<slug>.md. - Log it. One line in
_log.md, thenpython3 scripts/wiki.py all.
Contradictions get stated, not resolved
When two sources genuinely disagree and you cannot tell who is right, say so in both pages:
> CONTESTED: Source A says X (2026-03, link). Source B says the opposite (2026-07,
> link). Unresolved. Nothing here is measured, which is the actual finding.Quietly picking a side is how a wiki becomes confidently wrong. Naming the disagreement is more useful than a false answer, and it tells future-you exactly what to test.
Page format
---
topic: <one word>
updated: YYYY-MM-DD
sources: <count>
---
# Page title
The claim, up front. What is true and why it matters. Prose, not bullets — bullets let
you dodge the work of connecting two ideas.
## Connections
- [Other page](./other-page.md) — what the relationship actually is, not just "related"
## Sources
- [Author, "Title", date](../raw/YYYY-MM-DD-slug.md) — why this source is worth trustingThe Connections section is what makes this a wiki rather than a folder of notes. Say
what the relationship is. "Related" is not a connection.
Keeping it honest
python3 scripts/wiki.py hot # rebuild the cache
python3 scripts/wiki.py index # regenerate the index
python3 scripts/wiki.py lint # health check, exits 1 on findings
python3 scripts/wiki.py all # all threeThe lint catches broken cross-references, orphan pages nobody links to, citations that no longer resolve, and pages nobody has touched in months. It is mechanical — it cannot tell you two pages contradict each other. That is what the CONTESTED convention and an actual read are for.
Same principle as scripts/memory-lint.py, pointed at a different store: nothing will
ever prompt you to look. A wiki does not crash when it rots.
Boundaries
raw/is immutable. Capture it, never edit it.- One page per subject. Always extend, never duplicate.
- A source read but not written up did not happen. The reading is not the work.
- Delete a page that duplicates a live file elsewhere. Do not lint it harder. A clean check on the wrong artefact is worse than a dirty one.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

