Post

A Working Memory Vault for AI Coding Agents — Design and Practice of the Agent Working Memory Template

A Working Memory Vault for AI Coding Agents — Design and Practice of the Agent Working Memory Template

Introduction

AI coding agents are evolving fast. Claude now has Memory and session search. Context windows keep growing. Every generation is more capable than the last.

But progress creates a new problem.

Your knowledge becomes agent-dependent.

What you built up in Claude stays in Claude. Work done in Codex is locked inside Codex sessions. When a better agent comes along tomorrow — and it will — you start over. Re-explain your stack, your preferences, your decisions, your recurring patterns, from scratch.

This template exists for one reason:

To centralize your work knowledge in an agent-agnostic way, so no matter which agent you use — today or in the future — connecting the vault gives you immediate continuity.

Agents are tools. Tools change. But the knowledge and context you’ve accumulated should outlive any single tool.


What This Solves

1. Agent Lock-In

Claude Memory is useful inside Claude. Session search works within the same agent. Switch to Codex? Move to Cursor? Upgrade to the next-generation agent?

Knowledge locked inside a specific tool makes migrating to better tools expensive.

SituationProblem
Claude MemoryConvenient inside Claude, but no other agent can access it
Session searchSessions don’t evaporate, but they’re unstructured chat dumps
Agent migrationEvery new agent requires re-explaining context from scratch

2. Context Fragmentation

The more agents you use in parallel, the more fragmented your context becomes. “This decision was made in Claude, that code was written in Codex…” — everything scattered across different sessions.

3. Unstructured Information

Even with session search, chat logs are fundamentally unstructured. Decisions, work patterns, and reusable workflows are buried inside message streams, making them hard to reuse without manual extraction.


Design Principles

PrincipleCore Idea
Agent-AgnosticThe vault is not tied to any specific agent. Claude, Codex, or any future agent can read and write the same vault
CentralizedContext from every session converges in one place. No matter which agent you used, the information stays in the vault
Personal vault firstOptimized around one person’s work style, project status, and preferences. Not designed for multi-user sharing
Start with minimal routineMorning Briefing and Wrap-up are the default triggers. Tune timing and scope through actual use
Explicit sources & uncertaintyThe agent distinguishes files it actually checked from external sources it couldn’t access. Unqueried systems are marked “not queried”
Tiered recording levelsMust / Should / Maybe / Do not write — four levels to maintain signal-to-noise ratio
Bootabilityhot.md serves as a boot cache answering: “where are we now, and what should happen next?”

Overall Structure

The template consists of three layers.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Agent_Working_Memory_Template/
├─ AGENTS.md                         # Operating rules for the agent
├─ README.md                         # Quick-start guide for users
├─ SETUP.md                          # Setup questions and initialization guide
├─ hot.md                            # Boot cache for the next session
├─ log.md                            # Vault operation log (append-only)
├─ 00_Index.md                       # Vault index
├─ 01_Daily/                         # Daily notes and work logs
├─ 10_Projects/                      # Project wiki
├─ 20_Meetings/                      # Meeting notes and action items
├─ 30_Agent/
│  ├─ Working_Style.md               # User preferences and execution rules
│  ├─ Observations.md                # Recurring observed patterns
│  └─ Journal.md                     # Latest 1-2 session summaries
├─ 70_Entities/                      # People, tools, and concepts
│  ├─ People/
│  ├─ Tools/
│  └─ Concepts/
├─ 90_Retrospectives/                # Daily/session retrospectives
└─ 99_Templates/                     # Note templates

Three layers and their roles:

LayerRoleRepresentative Files
Agent ContinuityRestore the same context and style across sessionshot.md, Working_Style.md, Observations.md
Autonomous Wiki GrowthAccumulate work knowledge into a personal wiki10_Projects/, 20_Meetings/, 70_Entities/
Routine-Based RefreshPrevent the vault from going stale01_Daily/, 90_Retrospectives/, Journal.md

Core Routines

1. Setup (First-time configuration)

When the user says “Use this folder as my working memory vault,” the agent reads AGENTS.md and SETUP.md, asks 3–6 questions, and initializes the vault with user info, projects, briefing sources, and recording boundaries.

2. Morning Briefing

1
Please give me a morning briefing from this working memory vault.

The agent reads hot.md, the latest daily note, project index, and configured sources to summarize:

  • Today’s focus
  • Pending items
  • Blockers
  • External inputs
  • Independent work

If today’s daily note doesn’t exist, it’s auto-created from 99_Templates/daily.md.

3. During Work (Automatic recording)

The agent selectively records only what matters:

  • Must write — user corrections, explicit decisions, handoffs, external system status changes, meeting/document ingest results
  • Should write — recurring work patterns, reusable workflows, bug cause/fix patterns
  • Maybe write — brainstorming, one-off opinions, unverified interpretations
  • Do not write — simple conversational reactions, duplicates already in hot.md, unverified guesses, sensitive information

4. Wrap-Up

1
Please wrap up today.

The agent updates the daily note, creates/updates a retrospective, rewrites hot.md as a boot cache, and adds summaries to Journal.md and log.md. Recurring patterns are promoted to Observations.md or project notes.


Information Promotion Flow

1
2
3
4
5
6
Conversation
    → Daily note / Journal
    → hot.md
    → Project / Meeting / Entity notes
    → Observations
    → Working_Style
StageRole
ConversationRaw context needed for immediate work
Daily note / JournalCompressed record of the day or session
hot.mdShort boot cache for the next session
Project / Meeting / Entity notesLong-term work knowledge
ObservationsRecurring observed patterns
Working_StylePreferences validated enough to become execution rules

This flow prevents one-off information from becoming a permanent user rule too quickly.


This template does not replace Claude’s Memory or session search. It complements them.

FeatureStrengthLimitation
Claude MemoryFast in-agent reference, auto-learningClaude-only. No export/portability. Unstructured
Session SearchFind past conversationsUnstructured data. Hard to extract decisions/patterns. Agent-dependent
This VaultAgent-agnostic. Structured knowledge. Fully portableRequires initial setup. Needs routine maintenance

Using all three together is the most effective approach. Claude Memory for quick in-session reference, this vault for structured long-term preservation, and any agent can connect to it when needed.


Referenced Patterns

PatternCore IdeaReflection in Template
Karpathy LLM WikiA file-based wiki LLMs can read, organize, and connect over timeProjects, meetings, people, tools, and concepts are managed as Markdown wiki
OpenClaw Auto-DreamPeriodically summarize session history and promote important info to long-term memorySimplified into Journal → Observations → Working_Style flow
Daily Review / ShutdownUse daily reviews and shutdown rituals to define start and end of dayMorning Briefing and Wrap-up form the minimum refresh cycle

Expected Benefits

  • Use any agent — connect the same vault and start working immediately
  • Free migration — move from Claude to Codex to a future better agent without losing context
  • Decisions and corrections survive agent changes
  • Centralized context — even when using multiple agents in parallel
  • Knowledge scattered across chats accumulates into a personal work wiki
  • Claude Memory + vault = short-term reference + long-term preservation optimized
  • Morning Briefing and Wrap-up keep the vault from going stale
  • A human-readable file structure anyone can inspect, edit, and correct directly

This template is a personal working memory vault shaped around one person’s work style. It is not recommended as a shared team knowledge base.

  • Each person copies the template and uses it as a personal vault
  • The personal vault stores work style, current state, handoffs, and observations
  • Team-level decisions or project knowledge should be promoted to a separate team wiki or issue tracker
  • Sensitive information is not automatically recorded without explicit approval
  • Update timing and recording scope are adjusted through actual use

Useful Prompt Examples

IntentPrompt
Initial setup“Use this folder as my working memory vault. Please run setup.”
Start of day“Please give me a morning briefing.”
End of day“Please wrap up today.”
Meeting ingest“Please turn this meeting transcript into meeting notes and update the vault.”
Reflect a correction“The corrected schedule is important for the next session, so please reflect it in hot.md and the daily note.”
Exclude from record“This is personal, so please do not record it in the vault.”

Closing Thoughts — The Vault Outlives the Agent

Claude will keep evolving. Better agents will keep appearing. That doesn’t diminish the value of this template — it amplifies it.

The better agents get, the more important it becomes to keep your knowledge independent of any single one.

If tomorrow’s superior agent has no idea about the context you built in Claude today, it will have to learn everything about you from scratch. This vault bridges that gap. No matter which agent you use — today or five years from now — connecting the vault restores your work context instantly. That’s the ultimate purpose of this template.

GitHub: github.com/Lajancia/agent-working-memory-template

This post is licensed under CC BY 4.0 by the author.