Memory, session search, and skills for any AI agent.
Your coding agent remembers everything across every session — without a database, without a cloud service, without subscriptions. Just one npx command and SQLite.
npx mem-evolvedThree tiers of agent memory, same pattern that powers Hermes Agent:
| Tier | What | Why |
|---|---|---|
| Memory | Durable key-value facts | Your agent remembers preferences, conventions, and decisions across sessions |
| Session Search | FTS5 full-text search over past conversations | Nothing is lost — find any past decision, error, or context |
| Skills | Reusable markdown procedures | Knowledge compounds — your agent gets better at recurring tasks |
Auto-injection via MCP Resources — every time your agent starts, its stored memories are injected into the prompt automatically. No tool call needed. Your agent is smarter from turn 1.
# Run directly — no install required
npx mem-evolvedData is stored at ~/.mem-evolved/ — SQLite database + skills directory. Portable, inspectable, deletable.
Set a custom data directory:
MEM_EVOLVED_DIR=/path/to/data npx mem-evolvedAdd to ~/.claude/claude.json:
{
"mcpServers": {
"mem-evolved": {
"command": "npx",
"args": ["mem-evolved"]
}
}
}Restart Claude Code. Your agent now has persistent memory, session search, and skills.
Test it
> memory_add content="I prefer pnpm over npm" target="user"
> memory_list
> skill_save name="deploy-flow" content="# Deploy\n1. pnpm build\n2. pnpm deploy"
Add to .cursor/mcp.json in your project:
{
"mcpServers": {
"mem-evolved": {
"command": "npx",
"args": ["mem-evolved"]
}
}
}Open Cursor Settings → Features → MCP → Refresh. Tools appear automatically.
Test it
Ask your agent: "What do I prefer for package management?" after saving a memory. It will retrieve the fact from the auto-injected resource.
The pattern is the same — configure your MCP client to spawn npx mem-evolved as a stdio server.
| Tool | Args | Description |
|---|---|---|
memory_add |
content, target (user/memory) |
Save a durable fact |
memory_search |
query |
Full-text search across memories |
memory_list |
target (optional) |
List all stored memories |
memory_remove |
old_text |
Remove a memory by content match |
memory_replace |
old_text, content |
Update an outdated memory |
Target scopes:
"user"— personal preferences, style, habits (the agent's user)"memory"— project environment facts, conventions, decisions
| Tool | Args | Description |
|---|---|---|
session_search |
query, limit |
FTS5 full-text search over past sessions |
| Tool | Args | Description |
|---|---|---|
skill_save |
name, content, category |
Save a reusable procedure |
skill_load |
name |
Load a skill by name |
skill_list |
category |
List available skills |
skill_patch |
name, old_string, new_string |
Update a skill |
skill_delete |
name |
Remove a skill |
Three resources are exposed and read automatically by the agent every turn:
| URI | Content |
|---|---|
memory://current |
All memories, formatted for prompt injection |
memory://user |
User-specific memories only |
memory://project |
Project-specific memories only |
No manual tool call needed. The agent starts every session with context already loaded.
- Declarative only — save facts, not instructions. If a fact will be stale in 7 days, don't save it.
- Self-improving — skills patch themselves. If a skill is wrong, fix it immediately.
- Local first — all data in
~/.mem-evolved/. Your data stays yours. No cloud, no telemetry. - Zero config —
npx mem-evolvedis the only command. Everything else is your MCP client. - Portable — copy
~/.mem-evolved/to any machine. Resumes where you left off.
| Tool | mem-evolved |
|---|---|
| File-based context windows | Manual, fragile, easy to forget |
| Vector databases | Overkill for agent memory, need cloud or heavy infra |
| Hermes Agent | Full agent framework (task scheduling, TUI, plugins, 150+ tools) — mem-evolved is just the memory layer, extractable into any agent |
| In-memory conversation buffers | Lost on restart. mem-evolved persists across sessions. |
mem-evolved is the memory system extracted from Hermes Agent, packaged as a standalone MCP server. You get the same memory architecture without the full framework.
MIT