Every AI coding session starts with the same problem: the agent has no idea who you are, what you're building, or what you decided last week. You re-explain the stack. You paste context. You re-discover the same gotchas. Every. Single. Time.
Memex fixes this. It gives every AI agent — Claude, Cursor, Copilot, a local Ollama model, anything — persistent memory of your project. One session builds on the last. Decisions are remembered. Mistakes aren't repeated. Context follows the project, not the agent.
Day 1 Day 2
│ │
memex start claude memex resume claude
│ │
"Build me a checkout flow" "Continue from where we left off"
│ │
Claude works... Claude: "Yesterday we finished the
│ cart logic. Still need to wire up
You close the terminal Stripe webhooks. Want me to start?"
│ │
[Memex saves everything] Back in flow in 10 seconds.
Retains Claudes session
Sterne
4
Tools
—
Uptime
—
Switch tools mid-project without losing context. Start a feature in Claude CLI, continue it in Cursor, pick it up tomorrow with a free Ollama model. The memory lives in your project — not in any one tool.
Install
npm install -g @patravishek/memex
Memex uses a small AI call at the end of each session to compress the transcript into structured memory. You can use any of these — paid or free:
Provider
How to set up
Cost
Ollama (local)
ollama pull llama3.1 then export OLLAMA_BASE_URL=http://localhost:11434
# Example — pick whichever provider you're using (see Configuration below)
# Free: Groq
export LITELLM_BASE_URL=https://api.groq.com/openai/v1
export LITELLM_API_KEY=gsk_...
export LITELLM_MODEL=groq/llama-3.1-8b-instant
# OR: Anthropic (paid)
export ANTHROPIC_API_KEY=sk-ant-...
Then reload your shell:
source ~/.zshrc
That's the entire setup. See Configuration below for all supported providers including free options.
The two commands you need
memex start claude # Day 1 — or any time you start fresh on a project
memex resume claude # Every session after — Claude picks up where it left off
That's it. Everything else is automatic.
Using Cursor, Copilot, or another IDE agent instead of the CLI? Initialize without launching any agent:
memex init # creates .memex/memex.db and sets up the project
memex setup-mcp # connects Cursor Agent / Copilot via MCP
What it actually feels like
Day 1 — Starting a new project:
cd ~/my-project
memex start claude
You work with Claude for a couple of hours. Build some features, make decisions, hit a few problems. When you're done, close the terminal — any way you like (Ctrl+C, close the window, just kill it). Memex detects all of these and saves everything before the process fully exits.
Day 2 — Coming back:
cd ~/my-project
memex resume claude
Claude responds:
Continuing from our last session — we were building the checkout flow.
Here's what I have in memory:
• Project: my-app (Next.js + Stripe + PostgreSQL)
• Last session: Finished cart persistence, fixed guest cart bug
• Still pending: Wire up Stripe webhooks, add payment failure handling
• Watch out: Stripe webhook events can arrive out of order under load
What would you like to work on?
Claude is your personal engineering assistant — it remembers your project, your decisions, your mistakes, and exactly where you left off.
How it works
memex start claude
↓
Session recorded locally (.memex/sessions/)
↓
You exit → AI compresses transcript → .memex/memex.db
↓
memex resume claude
↓
Claude gets a short context hint + MCP tools connected
↓
Claude calls get_context(), get_gotchas()... on demand
↓
No wasted tokens. No performance warnings. Just context.
Memory is project-scoped — each project has its own independent database at .memex/memex.db. Nothing is shared between projects unless you choose to.
Install
Prerequisites
macOS
Node.js 18+
An API key from Anthropic, OpenAI, or a LiteLLM enterprise proxy
Memex picks the provider automatically based on what's set:
LiteLLM / Ollama / Groq / Gemini — if LITELLM_API_KEY + LITELLM_BASE_URL are both set
Anthropic — if ANTHROPIC_API_KEY is set
OpenAI — if OPENAI_API_KEY is set
All commands
memex init
Initialize Memex for a project without launching any agent. Creates .memex/memex.db and sets up .gitignore. Use this when working with IDE-based agents like Cursor or Copilot that don't need the CLI session wrapper.
cd ~/my-project
memex init
After running, open the Memex sidebar in VS Code / Cursor — or run memex setup-mcp to connect your IDE agent via MCP.
memex start [command]
Start a tracked session on a project for the first time.
memex start claude
memex start aider
memex start claude --project /path/to/project
On first run, Memex automatically adds .memex/ and .mcp.json to the project's .gitignore so session data is never accidentally committed.
When you exit — whether by typing exit, pressing Ctrl+C, or closing the terminal window — you'll see:
If the process is killed before compression finishes (e.g., the window is force-closed), Memex writes a recovery marker automatically. The next memex resume picks it up:
Recovering interrupted session from last run...
✔ Memory updated — focus: Implementing the checkout flow
Nothing is lost.
memex resume [command]
Resume with full context restored. Claude connects to Memex memory tools automatically on startup.
memex resume claude
Claude gets a brief summary of where things stand, then queries memory on demand — no 35k char dump, no performance warnings.
v0.4 options:
# Focus memory on a specific area — sorts gotchas, tasks, decisions by relevance
memex resume claude --focus "stripe webhooks"
# Control how much context is injected (1=one-liner, 2=key facts, 3=full)
memex resume claude --no-mcp --tier 2
# Hard cap on context size (~1 token = 4 chars)
memex resume claude --no-mcp --max-tokens 1000
# Combine: focus + budget
memex resume claude --no-mcp --focus "auth bug" --max-tokens 800
# Fall back to full context dump (no MCP)
memex resume claude --no-mcp
The --focus topic is saved to memory automatically — future resumes pick it up without you retyping it.
memex focus [topic]
View or set the current focus topic for this project.
# Set a new focus (saved to memory, used on next resume automatically)
memex focus "stripe payment integration"
# List current focus and full history
memex focus --list
# Clear focus entirely
memex focus --clear
Focus history keeps the last 10 topics. The AI also updates currentFocus naturally during end-of-session compression as work shifts — so history builds up even without explicitly running memex focus.
memex status
See what Memex currently knows about your project.
memex status
memex — project memory
Project: my-app
What this project does:
E-commerce platform with Next.js and Stripe
Tech stack: Next.js, TypeScript, Prisma, PostgreSQL
Current focus: Implementing the checkout flow
Pending tasks:
- Wire up Stripe webhooks
- Add payment failure handling
Gotchas:
- Stripe webhook events can arrive out of order under load
Sessions recorded: 4
Database: /my-app/.memex/memex.db
memex history
List past sessions with dates, durations, and one-line summaries.
memex history
memex history -n 50
#12 Feb 18, 2026 [claude] 42m
Finished cart persistence, fixed guest cart bug on page refresh.
#11 Feb 17, 2026 [claude] 1h 5m
Set up Stripe webhook handler and wrote integration tests.
memex show <id>
View the full transcript and summary of any past session.
memex show 12
memex search <query>
Full-text search across all past session summaries.
Make Memex memory tools permanently available in Claude without using memex resume.
# For this project only (commit to share with teammates)
memex setup-mcp
# For every Claude session on your machine
memex setup-mcp --global
After this, just run claude normally and say "resume from where we left off" — Claude will find the memory automatically.
memex compress
Manually re-run compression on the latest session. Useful if you want to force a re-summarize.
memex compress
In most cases you don't need this — if a session was interrupted (e.g., terminal force-closed), Memex automatically recovers it on the next memex resume.
memex prune [days]
Delete session records older than N days (default: 30).
memex prune # removes sessions older than 30 days
memex prune 7 # removes sessions older than 7 days
When Claude connects via MCP, it can call these tools on demand:
Tool
What it does
get_context()
Project summary, stack, current focus. Accepts optional focus (relevance sort) and tier (1/2/3 verbosity)
get_tasks()
Pending tasks
get_decisions()
Key architectural decisions with reasons
get_gotchas()
Pitfalls to avoid
get_important_files()
Files worth knowing about
get_recent_conversation()
Last N turns from previous session
search_sessions("query")
Search past session history
get_session(id)
Full detail of a specific session
save_observation(type, content)
Save a note, task, decision, or gotcha mid-session
save_observation is particularly useful — Claude can save important discoveries immediately without waiting for end-of-session compression.
Claude can also request a relevance-sorted view mid-session:
get_context(focus="payment flow", tier=2)
Privacy — <memex:skip> tag
Wrap any text in a session with <memex:skip>…</memex:skip> and Memex will strip it from the transcript before it's sent to the AI for compression. The raw session log is unaffected (local only), but the compressed memory will never contain that content.
Useful for passwords accidentally typed, personal context, internal URLs, or anything you don't want stored permanently:
<memex:skip>
Staging DB password: s3cr3t-temp
</memex:skip>
The AI sees [content excluded by <memex:skip>] in its place.
Project structure
.memex/
├── memex.db # SQLite database — memory + full session history
└── sessions/
├── *.jsonl # Structured session logs
└── *-raw.txt # Raw terminal recordings
.memex/ and .mcp.json are automatically added to the project's .gitignore on the first memex start or memex resume. Nothing needs to be done manually.
If .memex/ was already committed to a repo before upgrading to v0.4, run once:
Memory lives in .memex/memex.db — a local SQLite file that every agent reads from and writes to via MCP. Switch tools mid-project without losing a single byte of context.
Auto-configured via .cursor/mcp.json — install the VS Code extension or run memex setup-mcp
✅ Verified
OpenAI Codex
Auto-configured via .vscode/mcp.json — run memex setup-mcp
✅ Verified
GitHub Copilot
Auto-configured via .vscode/mcp.json — same extension
✅ Verified
Any MCP client
Run memex serve --project /path/to/project
✅ Works
Starter prompt for IDE agents:
When you open an IDE agent on a project, paste this once to restore context:
Please call get_context from the memex MCP to load project memory before we start.
After that the agent has your full project context — current focus, pending tasks, past decisions, gotchas — and you can continue where you left off.
The interoperability story (verified):
Day 1: memex start claude → work on project in Claude CLI
→ exit: memory compressed to .memex/memex.db
Day 2: Open Codex / Cursor → MCP connected automatically
"load context from memex"→ get_context() returns everything
→ continue exactly where Claude left off
No re-explaining. No copy-pasting. The context follows the project.
Notes
Claude CLI users: Memex automatically strips its own API keys from the environment before launching Claude, preventing the Auth conflict: Both a token and an API key are set warning.
Your data stays local: The only external call Memex makes is a single AI API call at the end of each session to compress the transcript. All memory, session logs, and MCP tool responses are purely local — no Memex server, no cloud, no account.
Enterprise: If your company uses LiteLLM, the compression call goes through your corporate proxy. The MCP tools make zero network calls — they read local SQLite.
Roadmap
Current: v0.4.5 — Crash recovery: sessions are never lost even if the terminal is force-closed
See ROADMAP.md for what's next — including a local web UI, semantic vector search, and more.
Contributions welcome. If a feature matters to you, open an issue.
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Fragen
Wie installiere ich Memex?
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
Benötigt Memex einen API-Schlüssel?
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
