PathruleField notes on
AI coding context.
Field notes on AI coding assistants, team memory, context engineering, token cost and security, with practical patterns for building consistent agent workflows.
- Engineering
What Survives a Claude Code Compaction?
After a compaction, Claude Code re-reads CLAUDE.md, unscoped rules and MEMORY.md from disk. What goes missing: decisions said in chat, path-scoped rules, long skills.
- Engineering
Why Not Just Let Your Coding Agent Grep?
Grep is the right way for a coding agent to find code. It cannot find what was never written there: decisions, reasons, lessons. What search misses and where to put it.
- Engineering
Where Do Coding Agents Save Their Own Memories?
Claude Code and Codex keep a private notebook on your machine, Copilot keeps repository facts on GitHub for 28 days, Gemini CLI dropped save_memory. Where seven agents write what they learn.
- Engineering
How Big Can AGENTS.md and CLAUDE.md Get?
Codex cuts AGENTS.md at 32 KiB, Claude Code skips a CLAUDE.md over 4 MiB, Windsurf caps rules at 12,000 characters. What seven coding agents actually read.
- Engineering
Introducing Pathrule Intelligence
An intelligence layer that learns how your workspace works and how you work from your own words, with no model on your machine, and carries both across coding agents. Why memory alone is not enough, why retrieval alone is not enough, and what we built instead.
- Engineering
Persistent Memory for AI Coding Agents: Six Decisions
Persistent memory for a coding agent is not one product choice. It is six: instruction file or layer, what a memory attaches to, how it is found, where it lives, when it arrives, and how you measure whether any of it worked.
- Engineering
Memory MCP Servers for AI Coding Agents: What Each One Stores
The protocol's reference server, Mem0, Zep through Graphiti, Supermemory, Basic Memory and what happened to OpenMemory. What each one calls a memory, where it runs, how to wire it into Claude Code, Codex, Cursor and OpenCode, and the one thing MCP does not settle.
- Engineering
AI Agent Memory: What It Is and How It Works
AI agent memory is the knowledge an agent can use in a session it did not learn it in. Here is the write path, the read path, the four kinds of memory and the blind spot every retrieval-only system has.
- Engineering
Best AI Agent Memory Tools for Coding in 2026
Most memory tools are not competing with each other. They store different units for different jobs, and picking the wrong category gives you a tool that works perfectly and still does not help.
- Engineering
How to Give Codex Persistent Project Memory
Codex supports local memories as well as AGENTS.md. Here is what persists, what teams can share, and where a cross-agent knowledge layer helps.
- Engineering
Vector RAG vs Path-Scoped Context for AI Coding
RAG is the right tool for a corpus you search. A codebase is not a corpus you search; it is a tree you are standing somewhere in, and the position you are standing in already tells you what matters.
- Engineering
How Do You Know the Fix Actually Worked?
Merging is a claim, not evidence. The runtime that reported the problem is the only thing that can say whether it stopped happening, and the honest answer has more than two states.
- Engineering
Pathrule vs Supermemory: A Memory API Answers When Asked
Supermemory is a good memory API and Pathrule is not competing to be one. The difference is delivery: retrieval waits for an agent to think of asking, and path-scoped context arrives before the first tool call.
- Engineering
What a Pathrule Pattern Is, and Why It Is Not a Skill
A pattern is not a single skill or rule. It is a small, opinionated bundle of memories, rules, and skills for one topic, each piece pre-scoped to the path it belongs to. Here is what that buys you and why the unit is bigger than a file.
- Security
Secure-by-Default Scaffolding for AI Coding Agents
A coding agent writes the first draft now, and the easy default is almost always the insecure one. Here is how a security pattern puts the secure default in front of the agent before it writes a single line.
- Engineering
One Import Token, Scoped Across Your Tree
Sharing conventions usually means pasting a long file at the root. A pattern is imported with one token that expands into scoped pieces, each placed at the path it applies to. Here is why that difference matters.
- Engineering
Why a Context Layer Should Have an Open Core
A local knowledge store, readable integrations and an honest source boundary. What Pathrule keeps open, and where proprietary Pathrule Cloud begins.
- Security
Where Your AI Context Layer Should Be Allowed to Run
For regulated and security-reviewed teams, the hard question about a context layer is not only what it stores. It is where it runs and who controls the data. Here is how to think about that line.
- Engineering
Context Engineering vs Prompt Engineering: What Actually Changed
Prompt engineering tunes the question. Context engineering decides what the model sees before it answers. Here is the difference, why it matters for coding agents, and how teams put it into practice.
- Engineering
What Is the Model Context Protocol (MCP) and Why Teams Use It
MCP is the open standard that lets AI assistants talk to external tools and data sources. Here is what it is, how clients and servers fit together, and where a team context layer plugs in.
- Engineering
Managing AI Coding Context in a Monorepo
A monorepo holds many projects under one root, which is exactly where a single AI instruction file breaks. Here is how to scope context per package so the assistant gets the right rules for the folder it is in.
- Engineering
GitHub Copilot Custom Instructions: What They Cover and Where They Stop
What Copilot’s repository and path-specific instructions handle, and where a shared knowledge layer adds versioned authoring and cross-agent delivery.
- Engineering
Why AI Coding Assistants Forget Your Codebase Between Sessions
AI coding assistants are stateless. Each new session starts from zero. Here is what actually persists, what does not, and how teams keep context across sessions.
- Engineering
How to Share Claude Code Memory Across a Team
Claude Code stores memory per project and per user by default. Here is how teams actually share that knowledge: what CLAUDE.md covers, where it stops, and what fills the gap.
- Security
Claude Code Hooks vs CLAUDE.md: Where Enforcement Lives
CLAUDE.md is advisory. Hooks are deterministic. Here is the practical split, why teams should use both, and where a path-scoped knowledge layer fits.
- Engineering
Why Claude Code Skills Do Not Activate
Skills look like they install cleanly and then quietly do nothing. Here is why SKILL.md often fails to activate and what a working delivery path looks like.
- Engineering
Why Cursor Rules Get Silently Ignored
Cursor rules look loaded and then quietly do nothing. Here is what breaks, what to check first, and how a path-scoped layer keeps rules visible at the moment they matter.
- Engineering
Why Claude Code Ignores CLAUDE.md
CLAUDE.md is guidance, not enforcement. Learn why AI coding assistants skip long instruction files and how hooks plus path-scoped rules close the gap.
- Engineering
AGENTS.md vs Pathrule: When Each One Makes Sense
AGENTS.md gives AI coding agents one predictable file. Pathrule keeps it short by routing the right team knowledge slice per session.
- Engineering
Context Is a Team Interface, Not a Prompt File
A practical argument for treating AI context as shared team infrastructure, not another prompt file to maintain by hand.
- Security
Why AI Coding Assistants Need Scoped Rules
Learn why AI coding rules work best when they are attached to the files and folders where they apply, reducing noise while preserving precise constraints.
- Token Cost
The Cost of Rediscovering Your Own Codebase
AI coding sessions often waste time and tokens rebuilding knowledge the team already has. Scoped context changes the first move.
- Engineering
From Tribal Knowledge to Reviewable Memory
Learn how teams turn scattered project knowledge into visible, scoped and reviewable memory that AI coding agents can reuse across sessions and tools.
- Security
A Practical Model for Safer AI Coding Workflows
Build safer AI coding workflows with explicit boundaries, scoped rules, reviewable memory, source-code privacy and clear human approval points for teams.
- Engineering
Why Global AI Instructions Break Down at Team Scale
Learn why global AI instruction files lose signal as teams add more paths and exceptions, and how scoped context keeps local constraints precise for teams.
- Engineering
Just-in-Time Context for AI Coding Agents
AI coding agents work better when relevant team knowledge arrives before the first useful action, not after a search loop.
- Engineering
Memory, Rules, and Skills: Three Kinds of Team Knowledge
Pathrule separates team knowledge into memories, rules, and skills so AI coding assistants receive the right kind of context.
- Security
How Stale Team Knowledge Creates AI Mistakes
Learn how stale team knowledge creates repeat AI coding mistakes, then use review, repair and retirement workflows to keep shared context trustworthy.
- Engineering
One Knowledge Layer Across AI Coding Agents
Teams should not rewrite the same guidance for every AI coding agent. Pathrule gives shared product knowledge one home across supported coding agents.
- Engineering
Design Rules Should Travel With the Work, Not the Tool
Keep AI-generated interfaces consistent by attaching product and design rules to the work itself, so the same constraints follow every coding tool you use.
- Engineering
Team Memory Should Not Change When AI Tools Do
Keep team memory stable while engineers switch AI coding tools, so shared decisions and conventions remain portable across agents, sessions and editors.
- Security
Audit Trails for AI Coding: What Changed, Who, and Why
An audit trail for AI coding shows who changed a rule, memory, or skill, when, and why. Here is why the context your assistants act on deserves the same review and history as code.
- Security
Pathrule Stores Team Knowledge, Not Source Code
Pathrule is built around a narrow privacy boundary: store the knowledge teams choose to write down, not their repository source.
- Engineering
Small Teams Should Build a Context Layer Early
Small teams feel AI context problems early. Capturing knowledge before the team grows keeps onboarding, review, and AI work calmer.
- Token Cost
More Context Is Not Always Better for AI Assistants
AI assistants do not need every fact at once. They need the right context, at the right depth, for the task in front of them.
- Case Studies
A Real History Test: 30x-50x Lower Context Cost
A real Pathrule history query showed how scoped prior work can avoid broad local search and reduce tool-output token cost.