You will re-explain your project structure to your agent. You will do it tomorrow, and the day after that, and the day after that. Every session starts from zero. The agent does not remember what you built yesterday, what decisions you made, or what it learned about your codebase. It just… doesn’t know.
This is the context window problem, and it is the single biggest friction point in working with AI coding agents today. Memory skills exist to fix it. The AgentNDX skills directory currently tracks six of them, and the category is growing fast.
What goes wrong without memory
An agent’s context window is both its workspace and its memory. Everything it knows about your project, your preferences, and the current task has to fit in that window. When the window fills up, older context gets compressed or dropped. When the session ends, everything is gone.
Three things break because of this.
First, session amnesia. You explain your coding conventions at 9am. By 9:01 tomorrow, the agent has no idea what you told it. Second, mid-session drift. Long sessions hit the context limit, and the agent quietly loses track of decisions it made an hour ago. Work quality drops and you might not notice until something breaks. Third — and this is the one that really stings — the agent never accumulates knowledge. Session 100 is the same blank slate as session 1. All that time spent teaching it your codebase? Gone.
Memory skills attack each of these problems differently.
Session state: surviving interruptions
Session State Manager saves agent progress at checkpoints during long tasks. If the context window resets or the session gets interrupted, the skill restores the agent to its last checkpoint instead of starting over.
The mechanics are simple. The skill watches for natural breakpoints in a multi-step workflow and writes a compact state file at each one: what has been done, what is in progress, what comes next, and any decisions made so far. When the agent resumes, it reads the state file and picks up where it left off. Useful for overnight runs, CI pipelines, and anything that outlasts a single context window.
Install: gh skill install sickn33/antigravity-awesome-skills/session-state-manager
Context management: fitting more in
Context Manager compresses conversation history without losing the decisions and constraints that matter.
Think of it as selective compression. The skill identifies which parts of the conversation contain actionable information — architecture decisions, user preferences, task requirements — and which parts are noise: exploratory questions, rejected approaches, verbose tool outputs. It keeps the signal, drops the noise. The result is that the agent can work through longer tasks without forgetting why it made a specific choice three hundred messages ago.
Install: gh skill install VoltAgent/awesome-agent-skills/context-manager
Long-term memory: remembering across sessions
This is where things get interesting.
Long-Term Memory from VoltAgent stores user preferences, past decisions, and project context in a persistent layer. At the start of each session, the skill retrieves relevant facts using semantic search. Your agent remembers that you prefer functional components over classes, that the API uses snake_case, and that the billing module is off-limits for refactoring. The first time it works, it feels like magic. Then it just feels like how things should have always worked.
Supermemory takes a similar approach but focuses on user profiling. It builds a structured profile of your preferences and working patterns over time, then injects a summary into the agent’s context at session start. The goal is to kill the “re-explaining” problem entirely.
Install: gh skill install VoltAgent/awesome-agent-skills/long-term-memory
Install: npm install supermemory
Structured memory: knowledge graphs and atomic notes
Two skills go beyond flat storage.
Knowledge Graph Builder extracts entities and relationships from documents and conversations, building a queryable graph. Instead of remembering raw text, the agent can query relationships: “which services depend on the auth module?” or “what changed in the billing API since last week?” If you work across a large codebase where relational context matters, this is the one to look at.
Ars Contexta takes a different approach — atomic notes with wiki-style linking. Every piece of knowledge gets stored as a small, self-contained note that links to related notes. Over time, the agent builds a connected web of knowledge about your project. It runs a processing pipeline that continuously distills and connects new information into the existing graph. Slower to set up, but the payoff compounds.
Both are in the pro tier on AgentNDX.
Memory skills vs. CLAUDE.md
If you use Claude Code, you already have a basic memory system: the CLAUDE.md file. It loads at the start of every session and gives the agent persistent instructions.
Memory skills are not a replacement for CLAUDE.md. They sit on top of it. CLAUDE.md handles static configuration — project rules, coding standards, tool preferences. Memory skills handle dynamic state: what happened last session, what the agent learned about your codebase, what decisions were made and why.
The two work together. Your CLAUDE.md tells the agent how to work. Memory skills tell it what it has already done.
Choosing the right memory skill
| Problem | Skill | Tier |
|---|---|---|
| Sessions end before the task is done | Session State Manager | Pro |
| Context window fills up mid-session | Context Manager | Free |
| Agent forgets everything between sessions | Long-Term Memory | Pro |
| Need structured, queryable project knowledge | Knowledge Graph Builder | Pro |
| Want zero-setup persistent user memory | Supermemory | Free |
| Building a connected knowledge base over time | Ars Contexta | Pro |
If you are just getting started, Context Manager and Supermemory are both free and solve the two most common problems. Start there.
FAQ
Do memory skills work with agents other than Claude Code? Most do. Context Manager and Supermemory both list Claude Code, Codex, Gemini CLI, and Cursor as compatible. Check the compatibility field on each skill’s directory page for specifics.
Do they slow down the agent? Not meaningfully. Reading a state file or injecting a memory summary adds a few seconds at session start. Mid-session compression runs in the background during natural pauses. For any session longer than 30 minutes, the tradeoff is worth it.
Can I use multiple memory skills together? Yes. A common setup is Context Manager for mid-session compression plus Long-Term Memory or Supermemory for cross-session persistence. They operate on different layers and don’t conflict.
Browse the full memory skills category on AgentNDX, or explore the complete skills directory for skills across all categories.