Mnemotron Wiki

A personal knowledge base that builds itself from your work.

Copyright (C) 2026 Patrick R. Wallace, Hamilton College LITS
Code: GPL-3.0-or-later | Documentation: GFDL-1.3-or-later | See License

AI-generated code notice: This software was developed with the assistance of Claude (Anthropic). All code has been reviewed and tested. Users are responsible for validating behavior in their own environments.

Mnemotron Wiki is a Claude-powered system that synthesizes professional documents, email, calendar data, and project notes into a structured, searchable Markdown wiki. You supply the raw material; Claude distills it into dossiers, topic pages, and a running index. Everything lives in plain Markdown files that you own, version with git, and read in any editor.

---

Contents

• How it works
• Privacy and data handling ← read before setup
• Requirements
• Quick start
• Detailed setup
• Usage guide
• Optional connectors
• Utility scripts reference
• Customizing the pipeline
• Scheduling automated runs
• Directory reference
• License
• Credits

---

How it works

The system has three layers.

1. Drop zone (raw_ingest/)
Drop any supported document here: PDFs, Word files, meeting notes, email exports. Optional connector tasks (Gmail, Google Calendar, Asana) can also write here automatically before a wiki update runs.

2. The wiki update task
Say "run the wiki update task" in a Claude session. Claude reads WIKI_UPDATE_TASK.md, scans raw_ingest/ for new files, synthesizes each document into wiki pages, updates the main index, and commits to git. A content-hash manifest ensures nothing is processed twice.

3. The wiki (wiki/)
Plain Markdown files in four directories:

Directory	Contents
wiki/daily/	Dated journal entries (from briefings or Asana summaries)
wiki/topics/	One page per project, initiative, or concept
wiki/people/	One dossier per person
wiki/organizations/	One page per org, team, or department

A typical run looks like this:

$ cd /path/to/my-wiki
$ claude
> Run the wiki update task.

[Claude scans raw_ingest/, processes 3 new files, updates 2 topic pages,
creates 1 new people page, regenerates INDEX.md, and commits.]

Wiki update complete — 2026-04-24
  Files processed:  3
  Pages updated:    2
  Pages created:    1
  Commit:           wiki update 2026-04-24

---

Privacy and data handling

Read this section before setting up the system.

What stays on your machine

Your wiki files are plain Markdown stored locally in a git repository. They do not leave your machine unless you push to a remote git repository or send them to Claude during processing.

What leaves your machine

When you run the wiki update task or any connector task, Claude reads the text of your source documents and writes results back to local files. The content you are processing — meeting notes, emails, calendar events, PDFs — is transmitted to Anthropic's servers during that processing session.

You are responsible for deciding what is appropriate to send. This system can be configured to process highly sensitive professional data. Consider the sensitivity of each category before enabling it.

---

Guidance by account type

Institutional accounts (enterprise agreements, BAAs)

If you are using Claude through an institutional agreement — a university, employer, or organization with a Business Associate Agreement or Data Processing Agreement with Anthropic — your institution's data handling policies govern what you may process. Before using this system with work data:

• Consult your IT security or legal team about what categories of data may be processed by an external AI service.
• Ask specifically about: employee communications, HR records, student records (FERPA in the US), financial data, health information (HIPAA), and legal documents.
• Verify whether your institution's agreement with Anthropic covers your specific use case (personal knowledge management vs. official workflows).

Claude.ai personal accounts (Free, Pro, Team plans)

By default, Anthropic may use content from conversations on personal plans to improve its models. If you are on a personal plan:

• Review Anthropic's current Privacy Policy and Terms of Service before processing any sensitive content. Policies change; always check the current version at anthropic.com.
• Enable privacy controls in your account settings. At the time of writing, claude.ai offers options to disable conversation history and opt out of model training — verify current options in your account.
• Do not process personnel records, salary or compensation data, medical or health information, legal documents under privilege, student records, or other regulated content through a personal plan without first understanding the privacy implications.
• Be aware that even with privacy settings enabled, Anthropic retains usage metadata (timestamps, token counts, feature usage).
• Consider whether the people you correspond with have consented to having their communications summarized by a third-party AI service.

Anthropic API (direct access)

Per Anthropic's API usage policy, content submitted via the API is not used to train models by default. However:

• Verify the current API data retention and usage terms at anthropic.com before processing sensitive data — policies may change.
• API-based Claude Code sessions have better data handling guarantees than personal claude.ai plans, but institutional review is still advisable for regulated content.

---

Special warnings for connector data

Gmail connector: Email is among the most sensitive data this system can process. Threads may contain personnel matters, compensation discussions, medical information, legal communications, and personal conversations. Use the filter rules in GMAIL_INGEST_TASK.md carefully. Err on the side of exclusion — if a thread would be uncomfortable to have summarized in a work document, it should not be processed.

Google Calendar connector: Calendar data reveals sensitive patterns: who you meet with, how often, and your schedule structure. The calendar connector is designed to capture structural information only (cadence, attendees, meeting series) and explicitly excludes meeting content. Review filter rules in CALENDAR_INGEST_TASK.md and exclude personal appointments.

---

Git remotes

If you push this repository to a remote (GitHub, GitLab, etc.):

• Use a private repository. Never push wiki content to a public repo.
• Your wiki content (dossiers, topic pages, daily entries) will be stored on the hosting service's servers.
• Consider whether your institution's policies permit storing work notes in personal or externally hosted repositories.

---

Requirements

Required:

• Python 3.10 or later
• Claude Code CLI, or claude.ai with Projects and file tools enabled
• An Anthropic account (institutional, personal, or API)
• Git

Optional (only needed for connectors):

• Gmail MCP connector configured in your Claude session
• Google Calendar MCP connector configured in your Claude session
• Asana MCP connector configured in your Claude session

---

Quick start

# 1. Clone or download
git clone <repo-url> my-wiki
cd my-wiki

# 2. Install Python dependencies
pip install -r scripts/requirements.txt

# 3. Personalize the starter files (see Detailed Setup)

# 4. Open a Claude session in the wiki root
claude

# 5. In your Claude session, say:
Run the wiki update task.

On first run with an empty raw_ingest/, the task scans, finds nothing to process, updates wiki/INDEX.md, and commits. That is the expected behavior — you are now ready to feed it content.

---

Detailed setup

Step 1: Clone or download

git clone <repo-url> my-wiki
cd my-wiki

If you downloaded a zip instead of cloning:

cd my-wiki
git init
git add .
git commit -m "initial mnemotron-wiki setup"

Step 2: Install Python dependencies

pip install -r scripts/requirements.txt

Package	Used for
pdfminer.six	PDF text extraction
python-docx	DOCX and ODT extraction
beautifulsoup4	HTML extraction
lxml	HTML parser (required by BeautifulSoup)
pyyaml	YAML frontmatter parsing (optional but recommended)

If you are not using Asana, you can skip pyyaml — the scripts include a minimal fallback parser for flat YAML frontmatter.

Step 3: Personalize the starter files

Your self-page: Rename wiki/people/your-name.md and fill in your own information. Claude uses this as context for understanding who you are and what you work on.

mv wiki/people/your-name.md wiki/people/alex-chen.md
# Edit the file to replace placeholder content with your own

Your organization: Edit wiki/organizations/lits.md to describe your institution or team. Replace all placeholder content.

The index: Edit wiki/INDEX.md to update the People and Organizations tables so they point to your renamed files.

Step 4: Open a Claude session in the wiki root

With Claude Code CLI:

cd /path/to/my-wiki
claude

With claude.ai Projects:

1. Create a new Project in claude.ai.
2. Enable file tools under Project settings.
3. Set the file tool's working directory to your wiki root.

Step 5: Run the first wiki update

In your Claude session:

Run the wiki update task.

Claude reads WIKI_UPDATE_TASK.md automatically — you do not need to paste it into the conversation. It will scan, find nothing to process on first run, regenerate the index, and commit.

---

Usage guide

Feeding content to the wiki

Drop any supported file into raw_ingest/ and run the wiki update task. Claude will extract the text, decide whether to create a new page or merge into an existing one, write the page, update the manifest, and delete the source file.

Supported file types:

Extension	Notes
.pdf	Text-layer PDFs only; scanned image PDFs extract minimal text
.docx / .odt	Word-compatible documents
.md / .txt	Plain text or Markdown
.html / .htm	Strips tags; extracts visible text
.csv	Reformatted as pipe-separated text for readability
.eml	Email export format; extracts headers and body text

Example workflow:

# Drop some files
cp ~/Downloads/q2-planning-notes.pdf raw_ingest/
cp ~/Downloads/annual-review-2025.docx raw_ingest/

# Check what is waiting to be processed
python scripts/check_ingest.py
# raw_ingest/q2-planning-notes.pdf
# raw_ingest/annual-review-2025.docx

# In your Claude session:
Run the wiki update task.

After the run, raw_ingest/ will be empty (files are deleted after successful processing) and the wiki will have new or updated pages.

What gets created from a planning notes PDF:

• A new wiki/topics/q2-planning.md (or an update to an existing page) with a summary, key points, and source document reference.
• New stub pages under wiki/people/ for any people named in the notes who do not yet have a dossier page.
• An updated wiki/INDEX.md.

People pages

Claude creates and updates people pages as names appear in source material. A typical page looks like this:

---
name: "Jordan Smith"
email: jsmith@example.edu
title: "Director of Digital Initiatives"
department: "Library"
organization: "Example University"
relationship: colleague
updated: 2026-04-15
---

# Jordan Smith

## Profile

Director of Digital Initiatives at Example University. Primary contact for
the shared metadata standards project.

## Personal Overview

Jordan communicates in terse, decision-focused emails — rarely more than
three sentences — but follows up quickly when action is needed. Tends to
front-load context in meeting agendas and prefers written summaries to
verbal debriefs.

## Working Relationship

Quarterly check-ins on the shared metadata project; occasional collaboration
on grant proposals.

## Interactions

| Date | Context |
|------|---------|
| 2026-04-10 | Metadata standards call — reviewed draft schema |
| 2026-03-22 | Intro meeting via DCT listserv |

The Personal Overview section is cumulative: it synthesizes communication style, working habits, and character from all available sources — documents, email, and calendar data. It grows richer over time.

You can add your own notes to any section. Claude preserves your edits and builds on them on subsequent runs.

Topic pages

Topic pages cover projects, initiatives, programs, or recurring themes:

---
updated: 2026-04-15
sources:
  - q2-planning-notes.pdf
tags:
  - planning
  - digital-collections
---

# Q2 Planning

## Summary

Quarterly planning session for Digital Collections. Team agreed to
prioritize the IA upload backlog with a target of 500 items by June 30.

## Key Points

The upload workflow bottleneck was identified as metadata review, not
the upload tooling itself. A dedicated metadata sprint is planned for May.
Three staff members will receive ArchivesSpace training before the sprint.

## Source Documents

| Document | Date Ingested | Notes |
|----------|--------------|-------|
| q2-planning-notes.pdf | 2026-04-15 | Q2 planning session notes |

Daily entries

If you use morning briefings (Markdown files dropped into raw_ingest/claude-briefings/ named morning-brief_YYYY-MM-DD.md), the wiki update task synthesizes them into dated daily entries at wiki/daily/YYYY-MM-DD.md.

---

Optional connectors

All connectors are independent. Use none, one, or all of them. Each writes to a subdirectory of raw_ingest/, which the wiki update task processes on the next run.

Gmail connector

Reads recent email threads and writes structured summaries to raw_ingest/gmail/. These become dossier updates for people pages and topic updates for project pages.

Setup:

1. Connect the Gmail MCP server in your Claude session:
   • Claude Code CLI: Add it to mcp_servers in your Claude Code settings (see Claude Code documentation for MCP configuration).
   • claude.ai: Settings → Integrations → Gmail.
2. In your Claude session:

   Run the Gmail ingest task.
   

3. Claude retrieves threads, filters noise (notifications, newsletters, automated alerts), clusters by person and topic, and writes summaries.
4. Run the wiki update task to incorporate the summaries into the wiki.

First run: Looks back 30 days. Subsequent runs look back 7 days.

What gets filtered: noreply senders, automated system notifications, calendar invites with no message body, mailing list digests, IT/HR system messages. Review and adjust the rules in GMAIL_INGEST_TASK.md Stage 3.

Privacy note: Review the Gmail warning above before enabling this connector.

---

Google Calendar connector

Reads recent and upcoming calendar events and writes structural summaries to raw_ingest/calendar/. These contribute to a meeting-cadence.md topic page and add attendance rows to people pages.

Important: The calendar connector captures structure only — who meets with whom and how often. It does not record what was discussed. Use the Gmail connector for meeting content.

Setup:

1. Connect the Google Calendar MCP connector in your Claude session (Claude Code mcp_servers config, or claude.ai Integrations).
2. In your Claude session:

   Run the calendar ingest task.
   

What it produces:

• A recurring-<slug>-<YYYYMMDD>.md file when the meeting cadence changes (new series, discontinued series, or changed attendees). The wiki update task synthesizes this into wiki/topics/meeting-cadence.md.
• An events-notable-<YYYYMMDD>.md file listing one-off significant meetings, upcoming commitments, and new contacts.

Lookback/lookahead: Defaults to 7 days past, 14 days ahead. Adjust in the Configuration table in CALENDAR_INGEST_TASK.md.

---

Asana connector

Ingests Asana project histories as topic pages. Asana files are idempotent refreshes of live project state: re-running the task updates the existing page rather than creating a duplicate.

How idempotency works:

The Asana ingest uses a separate manifest (.asana_manifest.json) keyed by project_gid rather than content hash. Claude can recognize the same project across multiple exports even when the content changes (updated task counts, new status messages, etc.).

Auto-generated content is wrapped in markers so user-authored notes are never overwritten:

<!-- asana:auto-start -->
... Asana snapshot and task history (regenerated each run) ...
<!-- asana:auto-end -->

## Notes
Your notes here are preserved across every refresh.

Setup:

1. Connect the Asana MCP connector in your Claude session.
2. Export a project history to raw_ingest/asana/ in the format described in WIKI_UPDATE_TASK.md Stage 3a. Asana files are identified by source: Asana and a project_gid in the frontmatter.
3. Run the wiki update task.

---

Utility scripts reference

Run all scripts from the wiki root directory.

scripts/check_ingest.py — list unprocessed files

python scripts/check_ingest.py              # files awaiting processing
python scripts/check_ingest.py --all        # all files including processed
python scripts/check_ingest.py --summary    # counts only

Example output:

raw_ingest/q2-planning-notes.pdf
raw_ingest/annual-review-2025.docx

Use this before a wiki update to see what is in the queue, or after to confirm everything was processed.

scripts/manifest.py — list processed files

python scripts/manifest.py

Example output:

Filename                                  Processed (UTC)            Wiki page
----------------------------------------------------------------------------------------------------
q2-planning-notes.pdf                     2026-04-15T14:32:00+00:00  wiki/topics/q2-planning.md
annual-review-2025.docx                   2026-04-15T14:35:00+00:00  wiki/topics/annual-review.md

scripts/extract_text.py — test text extraction

Verify a file will extract correctly before dropping it in raw_ingest/:

python scripts/extract_text.py path/to/document.pdf

Example output:

--- metadata ---
  title: Q2 Planning Session Notes
  author: Alex Chen
--- text ---
Q2 Planning — Digital Collections

Present: Alex Chen, Jordan Smith, ...
[full extracted text...]

If the command returns an error or empty text, the file will likely land in raw_ingest/failed/ during a real run. Common causes:

Problem	Cause	Fix
Empty output from a PDF	Scanned image PDF with no text layer	Run OCR first (e.g., ocrmypdf)
PasswordError	Encrypted PDF	Remove password protection before dropping
PackageNotFoundError	Missing dependency	Run pip install -r scripts/requirements.txt

scripts/asana_manifest.py — list Asana projects

python scripts/asana_manifest.py

Example output:

Project (gid)          Slug                                          Last ingested  Wiki page
----------------------------------------------------------------------------------------------------------
1234567890123456       dc-brothertown-project                        2026-04-15     wiki/topics/dc-brothertown-project.md

---

Customizing the pipeline

The *_TASK.md files are Claude's instructions. Edit them to change pipeline behavior — no code changes required. Claude reads the updated instructions on the next run.

Common customizations:

What to change	Where
What topics get their own pages	WIKI_UPDATE_TASK.md Stage 2
Topic page format and sections	WIKI_UPDATE_TASK.md Stage 2 (template)
People page format	WIKI_UPDATE_TASK.md Stage 2b (template)
Gmail filter rules	GMAIL_INGEST_TASK.md Stage 3
Gmail lookback window	GMAIL_INGEST_TASK.md Configuration table
Calendar lookback/lookahead windows	CALENDAR_INGEST_TASK.md Configuration table
Supported file extensions	SUPPORTED_EXTENSIONS in scripts/config.py
Adding a new file type extractor	scripts/extract_text.py — add a function and register it in _EXTRACTORS

---

Scheduling automated runs

Claude Code cron (recommended)

In your Claude session:

/cron add "30 8 * * 1-5" "Run the wiki update task."

To run Gmail ingest before the wiki update each weekday:

/cron add "15 8 * * 1-5" "Run the Gmail ingest task."
/cron add "30 8 * * 1-5" "Run the wiki update task."

System cron (Linux/macOS)

crontab -e

# Daily wiki update at 8:30 AM, weekdays
30 8 * * 1-5 cd /path/to/my-wiki && claude -p "Run the wiki update task."

launchd (macOS)

Create a .plist file in ~/Library/LaunchAgents/ to schedule via launchd. See Apple developer documentation for the .plist schema.

Suggested schedule

Task	Frequency	Notes
Wiki update	Daily (morning)	Processes whatever landed overnight
Gmail ingest	Daily, before wiki update	Run first; wiki update incorporates it
Calendar ingest	Weekly or as needed	Meeting cadence changes slowly
Asana ingest	Weekly	After significant project activity

---

Directory reference

mnemotron-wiki/
├── README.md                   ← you are here
├── LICENSE                     ← GPL-3.0 (code) / GFDL-1.3 (docs)
├── WIKI_UPDATE_TASK.md         ← main pipeline instructions for Claude
├── GMAIL_INGEST_TASK.md        ← Gmail connector instructions
├── CALENDAR_INGEST_TASK.md     ← Calendar connector instructions
├── .manifest.json              ← processed-file index (do not edit)
├── .asana_manifest.json        ← Asana project index (do not edit)
├── .gitignore
├── scripts/
│   ├── config.py               ← all path configuration
│   ├── manifest.py             ← content-hash manifest library
│   ├── check_ingest.py         ← list unprocessed files
│   ├── extract_text.py         ← extract text from PDF, DOCX, HTML, etc.
│   ├── asana_manifest.py       ← Asana project tracking (optional)
│   ├── ingest_asana_project.py ← Asana-specific ingest path (optional)
│   └── requirements.txt        ← Python dependencies
├── raw_ingest/
│   ├── claude-briefings/       ← morning briefing files
│   ├── asana/                  ← Asana project exports (optional)
│   ├── gmail/                  ← Gmail ingest output (optional)
│   ├── calendar/               ← Calendar ingest output (optional)
│   └── failed/                 ← extraction failures (quarantine)
└── wiki/
    ├── INDEX.md                ← auto-maintained table of contents
    ├── daily/                  ← dated journal entries
    ├── topics/                 ← project and concept pages
    ├── people/                 ← person dossiers
    └── organizations/          ← organization pages

Files managed automatically (do not edit manually)

Editing manifest files can cause the pipeline to reprocess already-ingested content or skip content it has not actually processed.

File	Managed by
.manifest.json	Wiki update task
.asana_manifest.json	Wiki update task (Asana path)
.calendar_manifest.json	Calendar connector
.gmail_manifest.json	Gmail connector

---

License

Code (GPL-3.0-or-later)

All Python scripts and supporting code in this repository are licensed under the GNU General Public License, version 3 or any later version.

Mnemotron Wiki
Copyright (C) 2026 Patrick R. Wallace, Hamilton College LITS

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.

Full license text: https://www.gnu.org/licenses/gpl-3.0.html

Documentation (GFDL-1.3-or-later)

All documentation in this repository — README.md, WIKI_UPDATE_TASK.md, GMAIL_INGEST_TASK.md, CALENDAR_INGEST_TASK.md — is licensed under the GNU Free Documentation License, Version 1.3 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts.

Full license text: https://www.gnu.org/licenses/fdl-1.3.html

AI-generated code notice

Portions of this software were developed with the assistance of Claude (Anthropic). All code has been reviewed and tested. Users are responsible for validating behavior in their own environments. No warranty is provided, express or implied.

---

Credits

Author: Patrick R. Wallace
Institution: Hamilton College Library and Information Technology Services (LITS)
Year: 2026

Mnemotron Wiki was developed at Hamilton College LITS as a personal knowledge management tool for AI-assisted institutional work environments.