Workplace Memory Architecture
"The difference between a brilliant new hire and an experienced colleague is not intelligence. It is context."
Zia Khan could hire the smartest analyst in Pakistan tomorrow. On day one, that analyst would ask: "What is Project Nighthawk?" "Who is Omar and what does he need from me?" "When you say Boulders, what do you mean?" The analyst is not slow — they are uninformed. They lack the accumulated context that every colleague builds over months of working together.
Claude faces the same constraint. Its intelligence is not the bottleneck. The bottleneck is organisational context — the structured knowledge of your people, your projects, your vocabulary, and your culture that every experienced colleague carries and that Claude, without memory, does not.
work.local.md encodes this context in four layers. This lesson explains each layer and builds Layers 1 and 4. Layers 2 and 3 follow in Lessons 4 and 5.
The Four Memory Layers
Each layer addresses a different dimension of organisational context. Together, they close all four failure modes introduced in Lesson 1.
| Layer | Name | What It Contains | Failure Mode It Closes |
|---|---|---|---|
| Layer 1 | Personal | Your working style, priorities, current focus, communication preferences | Priority Blindness (partial) |
| Layer 2 | Team | Key stakeholders — roles, communication styles, sensitivities, current focus | People Anonymity |
| Layer 3 | Projects | Active projects — status, owner, milestones, risks, decisions made | Project Amnesia |
| Layer 4 | Organisational | Terminology, meeting rhythm, culture, unwritten rules | Terminology Blindness |
The layers are designed to be built incrementally. Layer 1 comes first because it requires no information about others — just honest self-knowledge. Layer 4 comes second because the terminology dictionary has the most immediate, visible impact on output quality. Layers 2 and 3 build on this foundation.
How work.local.md Relates to CLAUDE.md
The official Productivity plugin creates CLAUDE.md — a hot cache that holds recent session context. It is optimised for speed: whatever Claude learned in recent sessions is stored here for fast retrieval.
work.local.md serves a different purpose. It is the structured professional memory layer — the enduring organisational context that does not change from session to session. Your terminology does not change weekly. Your key stakeholders do not change monthly. Your culture and unwritten rules do not change at all unless something significant shifts.
Both files exist. Both matter. They serve different functions in the same system.
CLAUDE.md: Hot cache — recent session context, temporary state
work.local.md: Professional memory — enduring organisational context
memory/ directory: Deep storage — longer-form reference documents
Layer 1: Personal Memory
Layer 1 is the calibration layer. It tells Claude who you are and how you work — so that every output is calibrated to your actual preferences, not to a generic professional.
The fields in Layer 1 are specific by design. Generic entries produce generic calibration.
Zia Khan's Layer 1:
## Layer 1: Personal
name: "Zia Khan"
role: "CEO, Panaversity / COO, PIAIC"
working_style: >
Direct; evidence-based; prefers bullet points for complex
information but narrative for persuasive content. Never
pad responses with preamble. Dense but precise.
decision_making: >
Systems thinker; wants to see second-order effects. Prefers
to understand the 'why' before the 'what'. Comfortable with
ambiguity but expects explicit flagging of key unknowns.
Evidence first, then judgment.
current_focus: >
AI Agent Factory book (completing Part 3 — Chapter 39 in
progress). PIAIC Karachi expansion (Project Nighthawk —
facility agreement stalled; needs escalation). Q2 2026
curriculum launch target.
working_hours: "07:00–22:00 PKT"
primary_tools:
- "Claude Cowork (primary AI interface)"
- "Google Workspace (calendar, docs, email)"
- "Notion (project management)"
communication_preference: >
Async-first; real-time for decisions and sensitive
conversations. Slack for daily; email for formal;
WhatsApp for urgent only.
language_preference: >
English for all professional content; technical terms in
English; greetings in Urdu acceptable.
Notice what this Layer 1 does that a generic profile does not: it specifies format preferences (bullets for analysis, narrative for persuasion), flags the current focus projects, and names the communication channels for different contexts. Every output Claude produces for Zia will be calibrated to these specifics.
The most important field — and the one most frequently underfilled — is current_focus. This field is the primary signal Claude uses to infer urgency. If AgentFactory is listed as the current focus, tasks connected to it receive elevated priority by default. Update this field whenever your focus shifts.
The template instructions say: "Work.local.md that describes how you wish you worked is useless. Describe how you actually work."
If you actually prefer to receive a summary before the detail — write that. If you actually check WhatsApp more than email — write that. If you actually make decisions quickly under pressure even though you prefer to have data — write that.
An aspirational Layer 1 calibrates Claude to a version of you that does not show up at work. The real you is what you need.
Layer 4: Organisational Memory
Layer 4 is where the Context Problem is most visibly solved. It contains four components — and the last one (unwritten rules) is consistently the highest-value, most underfilled section in every work.local.md.
The Terminology Dictionary
The terminology dictionary is the single most impactful change to AI output quality in this entire chapter. Before it exists, every output uses generic vocabulary. After it exists, Claude uses your organisation's actual language.
The format requires four pieces for each entry: the term, what it means, when to use it, and when not to use it.
Panaversity's terminology dictionary:
terminology:
"Boulders": >
Quarterly strategic priorities — our equivalent of OKRs.
Use: in all internal planning, status updates, and priority
discussions. Not: external communications or formal board
documents (use "strategic priorities" instead).
Related: "quarterly goals" in older documents.
"Digital FTE": >
A fully configured AI agent performing a specific professional
role. Not a chatbot — a Digital Full-Time Employee.
Use: all contexts where we discuss AI agents performing roles.
Not: marketing materials without explanation; new-hire onboarding
before the term is introduced.
"AgentFactory": >
Internal codename for the AI Agent Factory book and curriculum
project. Use: all internal references to the book project.
Not: external communications — use "AI Agent Factory" externally.
"Project Nighthawk": >
Internal codename for the Karachi expansion project.
Use: internal only — team discussions, planning documents,
status updates. Not: any external communication; do not use
externally under any circumstances.
"The Compass": >
The annual strategic planning document — reviewed quarterly.
Use: when referencing the strategic plan formally.
Not: casual references in Slack (use "the strategy doc").
"TutorClaw": >
The Digital FTE teaching agent for the AI Agent Factory
curriculum. Persona: knowledgeable guide; warm but precise.
Use: technical and curriculum contexts.
"PHM": >
Personalized Hybrid Model — Panaversity's seven-approach
adaptive teaching framework. Core pedagogical IP, owned by
Dr. Sana Mirza. Use: curriculum and pedagogical discussions.
"PIAIC Faisalabad SIG": >
AI Special Interest Group for business leaders in Faisalabad.
Zia chairs. Meets monthly. Use: when referencing this group
specifically. Not: confuse with the main PIAIC operations.
"CLEAR": >
Our values framework: Curious, Learner, Empathetic, Accountable,
Resilient. Use: when describing organisational culture or values-
based decisions. Not: as a casual adjective.
With this dictionary loaded, a request to "write a status update for the Boulders review" produces an output that uses "Boulders", references "AgentFactory" and "Project Nighthawk" by name, and refers to "Digital FTEs" correctly — without Zia needing to explain any of these terms.
Target for a well-built terminology dictionary: 20-50 entries. Start with terms that would confuse an outsider. Then add the terms that even colleagues within the organisation sometimes use inconsistently.
Meeting Rhythm
The meeting rhythm section teaches Claude the organisation's operational cadence. This is loaded automatically when preparing meeting briefs, the morning digest, and agenda planning.
meeting_rhythm:
- name: "Executive Weekly"
frequency: "Weekly"
day_time: "Monday 09:00 PKT"
attendees: ["CEO", "COO", "Heads of Department"]
purpose: "Weekly priorities; blockers; decisions needed"
format: "30 minutes standing; no slides"
note: "Hard stop at 09:30. Agenda items must be submitted
Friday afternoon."
- name: "Chapter Review"
frequency: "Bi-weekly"
day_time: "Friday 14:00 PKT"
attendees: ["Zia", "Publishing team"]
purpose: "Chapter draft review; feedback; next chapter planning"
format: "Review + feedback session"
note: "Chapters shared 48 hours in advance; feedback given
same day."
- name: "Banker Workshop"
frequency: "Monthly"
day_time: "Last Saturday of month — full day"
attendees: ["Zia", "Workshop team", "Banking sector attendees"]
purpose: "AI upskilling workshop for banking professionals"
note: "Content reviewed 7 days before delivery — non-negotiable."
Culture Configuration
The culture section — especially the unwritten rules — is where Layer 4 earns its investment.
culture:
values:
- "CLEAR: Curious, Learner, Empathetic, Accountable, Resilient"
decision_making: >
Evidence first; then judgment. Decisions documented in Notion.
"Disagree and commit" — once decided, full execution expected.
No revisiting decisions without new evidence.
communication: >
Async-first; real-time for decisions and sensitive conversations.
Slack for daily; email for formal; WhatsApp for urgent only.
Meeting invites sent with a minimum of 24 hours notice; same-day
invites require a direct message first.
unwritten_rules:
- "'Let's take this offline' = politically sensitive; do not
minute, do not reference in follow-up emails or shared docs"
- "Data requests to Omar need 3 business days' lead time
minimum; last-minute data requests are never acceptable"
- "Zia reviews book chapters personally before they go external;
do not share unreviewed drafts outside the team under any
circumstances"
- "Workshop content must be reviewed 7 days before delivery;
this deadline is non-negotiable regardless of other pressures"
The unwritten rules are the most valuable entries in the entire file. They encode the implicit norms that every colleague learns through experience — often after violating them once. A new hire who did not know that "let's take this offline" means "do not minute this" might write detailed meeting notes that include the politically sensitive discussion. An AI without this rule would do the same.
Every organisation has these rules. They feel obvious to insiders precisely because everyone knows them — but an AI cannot know them without being told. This section is where you tell it.
For every entry you consider adding to the unwritten rules section, ask: "If an AI assistant violated this rule in public output tomorrow — sent an email, produced a document, included something in a briefing — would it cause a problem?"
If yes, it belongs in the file.
Most professionals find they have between five and fifteen rules that pass this test. Write them all down.
The /agentic-office:workplace-context Skill
The workplace-context skill is the primary interface for reading from and writing to work.local.md. In this lesson, you are using it to verify that your Layer 1 and Layer 4 entries are producing the right outputs.
The skill handles six types of requests:
| Type | Example Trigger | What It Does |
|---|---|---|
| Verify layer content | "What do you know about how I work?" | Reads Layer 1 and summarises |
| Terminology query | "What does PHM mean?" | Looks up in Layer 4 dictionary |
| Add a person | "Add Dr. Sana Mirza to my team memory" | Formats entry for Layer 2 (Lesson 4) |
| Add a project | "Add Project Nighthawk to my projects" | Formats entry for Layer 3 (Lesson 5) |
| Describe organisation | "Describe our culture and working norms" | Reads Layer 4 and synthesises |
| Post-meeting update | "Update memory after today's meeting" | Proposes updates across all layers |
For this lesson, use the verification query:
/agentic-office:workplace-context
> Describe our organisation, our key terminology, and our
working culture using only what you know from work.local.md.
Sample output after Layer 1 and Layer 4 are populated:
WORKPLACE CONTEXT — Panaversity / PIAIC
════════════════════════════════════════════════════════════
PERSONAL CONTEXT (Layer 1)
You are Zia Khan, CEO of Panaversity and COO of PIAIC. Your
current focus is on three priorities: completing Part 3 of
the AI Agent Factory (Chapter 39 is in progress), resolving
the facility stall on Project Nighthawk, and hitting the Q2
2026 curriculum launch target.
Your preferred output format: bullet points for analysis;
narrative for persuasive content. No preamble. Evidence-
based recommendations.
ORGANISATIONAL CONTEXT (Layer 4)
Terminology:
Boulders → Quarterly strategic priorities (internal only)
Digital FTE → Fully configured AI agent in a professional role
AgentFactory→ Internal name for the AI Agent Factory book
Project Nighthawk → Karachi expansion (internal only)
The Compass → Annual strategic planning document
PHM → Personalized Hybrid Model (Dr. Sana Mirza's
framework; core pedagogical IP)
Meeting rhythm:
Monday 09:00 PKT → Executive Weekly (30 min standing, no slides)
Bi-weekly Friday 14:00 PKT → Chapter Review
Monthly last Saturday → Banker Workshop
Culture:
Values: CLEAR (Curious, Learner, Empathetic, Accountable,
Resilient)
Decisions: Evidence first; disagree and commit
Communication: Async-first; Slack daily / email formal /
WhatsApp urgent only
Unwritten rules loaded (4 entries)
LAYERS NOT YET POPULATED:
Layer 2 (Team): 0 people entries → Add in Lesson 4
Layer 3 (Projects): 0 project entries → Add in Lesson 5
════════════════════════════════════════════════════════════
This output demonstrates that the skill is reading correctly from work.local.md. It uses Panaversity's actual terminology. It references real projects by their internal names. It notes the meeting schedule. And — critically — it tells you what is still missing: Layers 2 and 3 are empty.
Compare this with what the same query produces on an empty work.local.md:
I don't have any organisational context loaded yet. To help
me act like a colleague rather than a generic assistant,
please share your work context using /agentic-office:setup
or tell me about your organisation, key people, and projects.
That is the before. The populated output above is the after. The difference between those two responses is the value of this lesson's exercise.
Exercise: Build Layers 1 and 4
This is the most important exercise in the chapter. A well-built Layer 1 and Layer 4 improves the quality of every skill lesson from here to Lesson 14. Do not rush it. Do not be aspirational. Be specific and honest.
Total time: 40 minutes
Part 1 — Build Layer 1: Personal (15 minutes)
Open your work.local.md file and fill in the Layer 1 section with your own real context.
Required fields:
nameandroleworking_style— how you actually communicate and process informationdecision_making— what you need before you make a decisioncurrent_focus— your top 1-3 priorities right now (update this frequently)working_hours— include timezoneprimary_tools— the two or three tools you use dailycommunication_preference— which channels for which purposes
Quality check: After filling in Layer 1, ask:
/agentic-office:workplace-context
> Based on my Layer 1, how should you format a project
status update for me? What should you never include?
If the answer matches your actual preferences — Layer 1 is correct. If it describes someone else, go back and be more specific.
Part 2 — Build Layer 4: Organisational Memory (25 minutes)
Layer 4 has four components. Work through them in order.
Step 1 — Terminology dictionary (15 minutes)
Aim for at least 10 entries to start. For each entry, include:
- What the term means in plain language
- When to use it
- When not to use it (especially: external vs. internal, formal vs. informal)
- Any related terms or synonyms that colleagues use
High-value entries to include first:
- Any project codenames (internal project names that differ from public names)
- Any methodology names your organisation has coined
- Any cultural phrases that have specific meanings ("Let's park this", "Let's align", "Let's socialise this")
- Any abbreviations that would confuse an outsider
- Any product, team, or programme names that are internal-only
Step 2 — Meeting rhythm (5 minutes)
Add your three to five most significant recurring meetings. Include the exact time and timezone — the digest skill uses this for meeting prep timing.
Step 3 — Culture and unwritten rules (5 minutes)
Add your organisation's stated values (use their exact wording). Write two to three sentences about how decisions are made and how information flows.
Then — the most important step — write the unwritten rules. Target: at least three rules. These should be things that every colleague knows but that are never in any handbook. Ask yourself: "What would an AI do wrong in its first week that would embarrass me?"
Verification Test
After completing both layers, run:
/agentic-office:workplace-context
> Describe our organisation, our key terminology, and our
working culture using only what you know from work.local.md.
Evaluation criteria:
- Does the output use at least three of your internal terms without prompting?
- Does the output reflect your actual working style preferences?
- Does the output mention your meeting rhythm?
- Does the output acknowledge what is still missing (Layers 2 and 3)?
If yes to all four — your foundation is solid. Lessons 4 and 5 build the people and project layers on top of this.
Your work.local.md is progressive — Lessons 4 and 5 build directly on what you create here. Do not start a new file or delete this one between lessons.
Use these prompts in Cowork or your preferred AI assistant.
Reproduce: Build a Layer 1 for Zia Khan using the case study data and test the output quality.
I am filling in work.local.md for Zia Khan, CEO of Panaversity.
Here is the Layer 1 data:
Name: Zia Khan
Role: CEO, Panaversity / COO, PIAIC
Working style: Direct; evidence-based; prefers bullet points
for complex information but narrative for persuasive content;
never pad responses with preamble
Decision making: Systems thinker; wants second-order effects;
prefers 'why' before 'what'; comfortable with ambiguity but
expects explicit flagging of unknowns
Current focus: AI Agent Factory Part 3 completion (Q2 2026
launch); Project Nighthawk facility escalation; BankersAI
Workshop 7 preparation
Working hours: 07:00-22:00 PKT
Tools: Claude Cowork, Google Workspace, Notion
Communication: Async-first; Slack daily, email formal,
WhatsApp urgent only
Based on this Layer 1, write a project status update for
Project Nighthawk. Do not include any information I have
not given you — only use this working context to calibrate
the format and style of the output.
What you are learning: Layer 1 does not provide project data — it calibrates format and style. The output should be structured correctly (bullets not paragraphs, direct not padded), but it will still ask you for the actual project status because that comes from Layer 3. This distinction — personal calibration vs. project data — is the difference between Layer 1 and Layer 3.
Adapt: Build your own Layer 1 and test how well it calibrates output.
I have filled in Layer 1 of my work.local.md. Here is my
working style entry:
[Paste your actual working_style entry here]
And my decision_making entry:
[Paste your actual decision_making entry here]
Write me a brief on a project called "[your current P1 project]"
in the format and style that these preferences suggest. Then
tell me: what did you change about your default output format
based on my Layer 1? What would you have done differently
without it?
What you are learning: The "what would you have done differently" question is diagnostic. If the answer is "not much" — your Layer 1 entries are too generic. Sharpen them. If the answer identifies three or four specific adjustments — your Layer 1 is working.
Apply: Build a complete Layer 4 terminology dictionary for your organisation and test it.
Here are 10 internal terms my organisation uses. For each,
I will give you the term and its real meaning. Format each
as a work.local.md terminology entry (definition, when to
use, when NOT to use, related terms):
1. [Your term] — means: [definition]
2. [Your term] — means: [definition]
3. [Your term] — means: [definition]
[continue to 10]
After formatting these as proper terminology entries, write
a status update for one of my projects using ONLY these
internal terms — no generic substitutes. Then write the
same update as an external stakeholder communication, applying
the "Not:" guidance from the terminology entries to avoid
internal-only language.
What you are learning: The "internal vs external" version of the same update demonstrates the value of the "Not:" field in the terminology dictionary. Claude applying the "Not: external communications" guidance automatically — without you needing to say it — is exactly the behaviour the terminology dictionary enables.
Flashcards Study Aid
Continue to Lesson 4: Building Your People Memory →