Every AI agent has the same blind spot: it forgets. A fresh context window means a fresh start. The decisions you made, the constraints you set, the bugs you already debugged — gone. Memory skills exist to fix this. They give agents the ability to store, compress, and retrieve context across sessions, so work compounds instead of resetting.

These five skills cover different approaches to the memory problem. Some focus on within-session context compression. Others handle cross-session persistence. A few build structured knowledge that agents can query later. All run inside agents like Claude Code, Codex, and Cursor.

What to look for

The first question is scope: do you need memory within a single long session, or across multiple sessions? Within-session skills focus on compression and retrieval — keeping relevant context alive when the token window fills up. Cross-session skills write to disk or a database and retrieve on demand.

Storage format matters more than you’d expect. Flat key-value stores work for simple preferences. Knowledge graphs capture relationships between entities and decisions. Atomic note systems let agents build and traverse linked references. Pick the format that matches how your agent thinks about its work.

Retrieval quality separates good memory skills from bad ones. Storing everything is easy. Knowing what to surface when — that’s the hard part. Look for skills that support semantic search, relevance scoring, or structured queries instead of brute-force context dumps.

Finally, check compatibility. Memory skills need tight integration with your agent’s session lifecycle. A skill that saves state but can’t hook into your agent’s startup sequence is half a solution.

Top agent skills for memory and context management

1. Context Manager

Context Manager handles the most common memory problem: long sessions that overflow the context window. It compresses, indexes, and retrieves conversation history so agents can run multi-step tasks without losing track of earlier decisions.

What sets it apart is the compression strategy. Rather than dropping old messages wholesale, it indexes key decisions and constraints, then retrieves them when they become relevant again. You get the effect of a much larger context window without the token cost.

Wide compatibility helps too. It works with Claude Code, Codex, Gemini CLI, Cursor, and most universal agent setups.

Compatible with: Claude Code, Codex, Gemini CLI, Cursor, Universal Category: Memory Install: gh skill install VoltAgent/awesome-agent-skills/context-manager

2. Long-Term Memory

Long-Term Memory solves the cross-session problem. It stores facts, preferences, and decisions in a structured key-value format with semantic search, so agents can recall relevant context from past interactions without you repeating yourself.

The retrieval layer is the real value. It doesn’t dump everything from storage — it finds the entries that match the current task using semantic similarity. That means an agent picking up a project from two weeks ago can surface the relevant constraints without loading its entire history.

Useful for any workflow where agents touch the same project repeatedly: ongoing codebases, recurring reports, client work with accumulated context.

Compatible with: Claude Code, Codex, Universal Category: Memory Install: gh skill install VoltAgent/awesome-agent-skills/long-term-memory

3. Session State Manager

Session State Manager takes a checkpoint-based approach. It saves agent progress at key moments — before a context window reset, at the end of a subtask, or when you explicitly ask — and restores that state when the agent picks up again.

Think of it like save points in a game. The agent captures its current task, open questions, completed steps, and pending decisions. On resumption, it loads the checkpoint and picks up where it left off. No manual re-prompting required.

Particularly useful for long-running tasks that span multiple sessions: migrations, refactors, multi-file features, or research that takes days.

Compatible with: Claude Code, Codex, Gemini CLI, Universal Category: Memory Install: gh skill install sickn33/antigravity-awesome-skills/session-state-manager

4. Knowledge Graph Builder

Knowledge Graph Builder takes a more structured approach to memory. Instead of storing flat text, it extracts entities and relationships from documents or conversations and builds a queryable graph. Agents can then traverse relationships, not just search keywords.

The graph structure pays off when your agent needs relational context — “what services depend on this database?” or “which decisions were influenced by the Q1 performance review?” Flat retrieval can’t answer those questions. Graph traversal can.

Works well as part of a RAG pipeline. Feed it your documentation, meeting notes, or codebase context, and it builds the relationship map your agent can query during tasks.

Compatible with: Claude Code, Codex, Universal Category: Memory Install: gh skill install VoltAgent/awesome-agent-skills/knowledge-graph-builder

5. Ars Contexta

Ars Contexta is the most opinionated skill on this list. It builds a persistent skill graph using atomic Markdown notes, wiki-links, and Maps of Content — a system inspired by Zettelkasten and adapted for agents. Three-space architecture: self (identity), notes (knowledge), and ops (workflows).

The approach solves the blank-slate problem differently than the others. Instead of storing raw session data, it processes knowledge through a 6R pipeline: read, reduce, relate, record, review, refine. The result is a structured, linked knowledge base that grows over time and survives context resets.

Best for agents that run daily or weekly on the same domain — where accumulated, interconnected knowledge matters more than raw recall.

Compatible with: Claude Code, Universal Category: Memory Install: gh skill install agenticnotetaking/arscontexta

How to choose

If your main problem is sessions running out of context, start with Context Manager. It handles compression and retrieval within a single session and works with the widest range of agents.

If you need cross-session persistence — remembering what happened last Tuesday — Long-Term Memory or Session State Manager are your picks. Long-Term Memory is better for factual recall (preferences, decisions, constraints). Session State Manager is better for task continuity (picking up where you left off).

For structured, relational knowledge that agents can traverse, Knowledge Graph Builder is the right choice. It’s heavier to set up but pays off when your agent needs to reason about relationships, not just retrieve facts.

And if you want a full knowledge system that your agent maintains and grows over time, Ars Contexta is the most complete framework — though it requires buying into its opinionated structure.

Most production setups combine two: a within-session skill (Context Manager) plus a cross-session one (Long-Term Memory or Session State Manager). That covers both the “don’t forget mid-task” and “don’t forget between tasks” failure modes.

FAQ

Q: Can I use multiple memory skills at the same time? A: Yes, and many setups do. A common pairing is Context Manager for within-session compression plus Long-Term Memory for cross-session persistence. They handle different failure modes and don’t conflict.

Q: Do memory skills slow down my agent? A: Retrieval adds a small overhead per turn — usually a few hundred milliseconds for semantic search. Context Manager and Session State Manager run inline and rarely cause noticeable delay. Knowledge Graph Builder can be slower during the initial graph construction phase, but queries are fast once built.

Q: How much storage do these skills need? A: Most use local Markdown files or lightweight SQLite databases. Even after months of use, storage is measured in megabytes, not gigabytes. Knowledge Graph Builder and Ars Contexta produce more files but still stay small relative to the codebases they help manage.