MemCell MCP server

memcell

Persistent, adaptive memory engine for AI agents. Stop your agents from repeating the same mistakes across sessions.

= 20" />

# Install globally
npm install -g memcell

# In any project:
memcell connect

Run it in any project. memcell connect automatically:

• Approves in 1 tap: Opens browser approval (no upfront account needed; guest memories carry over seamlessly on sign-in).
• Auto-detects agents: Wires Claude Code, Cursor, Copilot CLI, Gemini CLI, OpenAI Codex, OpenCode, Cline, Windsurf, and more with non-destructive lifecycle hooks and the local stdio MCP server (memcell mcp).
• Secures credentials: Stores scoped agent keys directly in your OS keyring (Keychain / Secret Service / Credential Manager).
• Creates .memcell: Adds a clean, git-safe project descriptor to your repo root.

Verify anytime:

memcell status

Why MemCell?

Every time you start a new session with an AI agent, it starts with complete amnesia. It falls into the same traps, repeats outdated assumptions, and recreates issues you already resolved yesterday.

Static instruction files and monolithic system prompts don't solve this:

• Context Bloat: Stacking dozens of instructions wastes tokens on every turn and dilutes model attention.
• Zero Adaptability: Outdated instructions keep firing even when requirements change.
• Manual Overhead: You remain the sole feedback loop, constantly hand-editing markdown files.

MemCell gives your agents closed-loop memory:

1. Pre-Action Recall — Before taking action, agents receive verified statements relevant to the current task.
2. Execution & Feedback — When actions succeed or fail, real outcomes are reported back.
3. Earned Confidence — Statements that work gain confidence; statements that fail decay. Zero manual prompt maintenance.

How It Works

MemCell pairs two lightweight layers so agents stay aligned without getting in your way:

• Deterministic Hooks (Zero-Touch): Fire automatically on session start, prompt submit, and turn end. Relevant statements and directives are checked before actions begin—without relying on the model remembering to call a tool. Hooks fail open in <50ms, so offline or slow networks never block your editor.
• MCP Tool Bridge (memcell mcp): Provides interactive tools (recall, remember, report, ingest) when agents need to search memory or record discoveries mid-turn.

Everyday Usage

Once connected, agents recall and report memory automatically. You can also interact with memory directly from your terminal:

1. Recall Verified Context Before Implementing

memcell recall "how do we handle multi-tenant database sessions?"

2. Record a Decision or Convention

# Record a project convention
memcell remember "Always use pnpm for package operations; never invoke npm directly"

# Bind an invariant to a specific lifecycle trigger
memcell remember "Check database migrations for column locks before altering tables" --at "before db:migrate"

3. Report What Happened

Memories sharpen when told what happened. When an approach succeeds or fails, report the outcome:

Shortened. The full README is on GitHub.