agent-memory
Installation
SKILL.md
Agent Memory
Overview
Coding agents start every session with an empty context window. This skill gives them memory that survives: first the file-based memory the agents themselves read (CLAUDE.md, AGENTS.md, GEMINI.md), then a small memory module you own (markdown files, SQLite with embeddings, or ChromaDB) for what instruction files cannot hold, such as thousands of past tickets or conversations searched by meaning.
Instructions
Step 0: Use the memory your agent already has
Check this before building anything. Instruction files are loaded at the start of every session:
| Agent | Persistent instructions | Notes |
|---|---|---|
| Claude Code | CLAUDE.md (project, ~/.claude/CLAUDE.md), CLAUDE.local.md, .claude/rules/*.md |
Imports with @docs/api.md. Auto memory is on by default: Claude writes notes to ~/.claude/projects/<project>/memory/, MEMORY.md is loaded at start (first 200 lines or 25KB). Browse or toggle with /memory; disable with CLAUDE_CODE_DISABLE_AUTO_MEMORY=1. |
| OpenAI Codex | AGENTS.md (~/.codex and each directory from the repo root down), AGENTS.override.md |
Combined size capped by project_doc_max_bytes (32 KiB). |
| Gemini CLI | GEMINI.md (~/.gemini/ and project directories) |
/memory show and /memory reload; the file name can be changed with context.fileName (for example AGENTS.md). |
If the request is "make the agent remember our conventions", write or extend those files and stop. Build the strategies below only for memory that is too large, too dynamic or must be searched semantically.