Skip to main content

Working Memory

How many things are you trying to hold in your head right now?

JugglingWriting it down
Forget where you wereKnow where you left off
Shallow work all dayDeep work by design
Drag from the last thingClean start every time

Working memory is volatile. Task switching can evict the state needed for the current decision and force you to reconstruct it later.

Treat five things as a local interface budget, not a universal biological law. The useful rule is simpler: externalize volatile state, then retrieve only the context needed for the decision in front of you.

Flow State

Overloading and needless Context Switching is the enemy of flowing progress.

Your brain is for having ideas, not holding them.

Deep work needs a clear head. A clear head needs a place to put things. That place needs a rhythm — morning to load, day to capture, evening to clear. A text file does this. Everything else exists to protect the conditions for deep work.

Four Types

The same types show up in people and machines:

TypeYouA MachineWhat to Do
WorkingThe active decision stateContext windowExternalize volatile state
EpisodicWhat happened to youLogs, conversation historyJournal it while fresh
ProceduralSkills you practisedTrained behavioursRepeat until automatic
SemanticFacts you knowKnowledge basesLook it up, don't memorise it

Better memory management means better thinking. True for people. True for machines.

Get It Out

Put everything into a system you trust:

WhatWhereWhy
TasksOne task listUnwritten tasks nag in the background
DecisionsShared notes, right afterYou forget the details within hours
IdeasA capture tool you always haveGood ideas rot fast
Context"Where I left off" notesSaves twenty minutes tomorrow morning
FactsA searchable placeNever memorise what you can look up

The test: can you forget about it and trust the system to hold it? If not, your head is still carrying the weight.

Five as a Design Budget

Five is a useful compression constraint for the Tight Five and small decision interfaces. It is not a claim that every person, task, or machine has exactly five fixed slots.

Try fourWhat falls out
Drop PerformanceNo scoreboard — you drift without knowing
Drop PlayersNo ecosystem — you build alone without calibration
Try sixWhat blurs
Add "Culture"Already inside every slot — not a separate thing
Split Platform in twoSame lens, same slot — you just can't see it yet

Four loses a dimension. Six hides a duplicate. Five is the minimum viable picture of the whole. The Tight Five is not a framework you adopt. It is a constraint you discover.

The same five recurs here because a stable local pattern reduces retrieval and comparison cost:

ScaleThe Five
Working memoryFive slots in your head
Tight Five — innerPerformance, Principles, Platform, Process, Players — what you control
Tight Five — outerPurpose, Principles, Platform, Perspective, Performance — how you serve
Business hatsStrategy, Production, Technology, Finance, Compliance
Customer journeyAttraction, Persuasion, Navigation, Validation, Renewal
PromptDeckFive slides, five questions, five answers
Work frontsStrategy, Project Mgt, Marketing, Platform, Performance
Week planFive bets max. Five days. Five lenses per session.

Same pattern, different zoom. If five hides a needed distinction, change the local interface rather than treating the number as doctrine.

Chunking

Five slots. Make each slot hold more:

ScatteredChunked
Eight random digitsFour two-digit pairs
Twelve loose tasksThree projects, four tasks each
A long argumentSetup, core point, conclusion

Expertise changes how information is chunked and retrieved. The design goal is not to maximize what stays in mind; it is to make the next relevant chunk cheap to recover.

Protect It

ThreatWhat it costsDefence
NotificationsTwenty minutes per interruptionSilent mode during deep work
Switching tasksNearly half your productive timeOne theme per day
Too many tabs openAttention split across all of themClose what you don't need now
Back-to-back meetingsResidue from the last oneFive-minute gap between them
No clear next stepYou stall and driftWrite the next step before you stop

The Text File

A plain text file. One section per day. Three parts. Nothing else.

## 2026-03-25

### Plan
- [ ] The work that fits today
- [ ] One more thing, maybe two

### Capture
- thing that came up during work
- idea that arrived uninvited

### Clarify
- Done: what got finished
- Moved to Friday: what didn't
- Killed: what turned out not to matter

The rules:

  1. Nothing carries over. If it didn't happen today, give it a future date or kill it.
  2. Plan the night before. Your sleeping brain loads context overnight.
  3. One theme per day. Switching themes costs almost half your output.
  4. Capture without stopping. When something comes up, write it down and keep going.
  5. Clarify every evening. Move it, finish it, or kill it. Nothing stays in capture overnight.
  6. Five to seven items. More than that means you haven't grouped them.
  7. Keep old days. The file grows down. You can search it later.

Plan loads the buffer. Capture keeps it from overflowing. Clarify empties it. The file is today's workspace, not a permanent record.

Human and Machine

The same file works for both:

YouYour AI agent
PlanMorning — what matters todaySession start — what to work on
CaptureThings that come upDiscoveries mid-session
ClarifyEvening — move it or kill itSession end — update the log or kill it
HandoffYour "where I left off" is the agent's morning briefingThe agent's summary is your context next session

Your evening writes the agent's morning. The agent's close writes your next start. The text file is where the two meet.

The Day

  1. Morning — Read your file. Pick the one thing.
  2. Deep work — Best hours. One task. No interruptions.
  3. Admin — Inbox, replies, updates. Batch it.
  4. Close — Write where you left off. Clear the capture list. Shut down.
  5. Rest — Trust the system. Let your brain recover.

The Trap

Collecting instead of doing. Refusing to write it down because you think you should remember it. Perfecting the system instead of using it. Organising as a substitute for output.

The system exists to free your mind for work. If you spend more time on the system than on the work, the system won.

By Archetype

ArchetypeHow they use memory
EngineerHolds complex systems — needs external scaffolding
PhilosopherConnects across domains — needs long unbroken stretches
CoachRemembers people — needs context about relationships

When To Use

Use this pattern the moment your day fragments across too many open threads. The principle is one trusted place to put everything, cleared on a fixed rhythm. Reach for it when you notice yourself re-deriving where you left off, or when an agent session reboots with no memory of the last one.

Use the Knowledge Schema for durable memory: one canonical home, stable chunks, progressive disclosure, retrieval triggers, and typed relationships. Working memory should carry only the current decision and the evidence it needs.

Retrieval trigger: return when a person or agent must reconstruct the active decision after a task switch. Store durable context in its canonical Playbook or governed project home; keep only volatile session state in the daily buffer. Evidence of success is less re-derivation without a larger active context.

Failure Modes

  • The system becomes the work. Perfecting the file instead of clearing it. The rhythm exists to free the mind, not to be tuned.
  • Backlog creep. Letting yesterday's uncleared items ride into today. Nothing carries over — date it forward or kill it.
  • Overloading a slot. Cramming eight threads into five. When the picture blurs, group before you add.
  • Silent staleness. Trusting a buffer that has gone cold. If you cannot state the next step from the file, the buffer failed.

Context

Signal

Changes my mind: show that externalizing and retrieving decision state adds more cost than it saves across repeated task switches; the case for the buffer would narrow.

Next question: what is the smallest shared file format that a human and an agent can both write to and read from without translation?

Questions

What happens when a person and a machine share the same memory file?

  • What is the smallest file that supports both a human day and an agent session?
  • When you write "where I left off" and a machine reads it tomorrow, what survives and what gets lost?
  • Which chunk and retrieval cue let an expert recover the needed model with the least noise?
  • How do you know your buffer is stale before it costs you?