The biggest problem with AI agents (Cursor, Claude Code, OpenClaw) is that every session starts from zero.
You explain the architecture, it does great work, and in the next session — it forgot everything.
All we have today are MEMORY.md and CLAUDE.md files that max out at 200 lines and go stale.
What is MCP?
Model Context Protocol is an open standard created by Anthropic.
The idea is simple: instead of every AI agent saving memory in its own format, there's one unified protocol they all speak.
An MCP server runs in the background and provides tools to any agent that connects — search, save, retrieve.
AgentMemory — The Implementation
AgentMemory is an MCP server that adds persistent memory to any AI agent.
It runs locally on your server (self-hosted, zero cloud dependency) and provides:
- Triple-stream search — BM25 (exact keywords) + Vector (semantic meaning) + Knowledge Graph (relationships)
- 95.2% retrieval accuracy on ICLR 2025 benchmark
- ~80% token savings — instead of injecting all context, injects only what's relevant
- Local embeddings —
all-MiniLM-L6-v2, no API key, no cost
- Auto-forget — automatically cleans irrelevant memories
- 51 MCP tools — search, save, timeline, profiles, and more
How it works in practice:
# AgentMemory server runs in the background
npx @agentmemory/agentmemory
# AI agent connects via MCP
# Session 1: you set up JWT auth with jose
# Session 2: you ask for rate limiting
# → The agent already knows your auth is in src/middleware/auth.ts
# and that you chose jose for Edge compatibility
# Zero repeated explanations.
Why does this matter?
The context window problem isn't just size — it's cost.
Opus 4 costs $15/M input tokens. If you inject 20K tokens of context every message,
that's ~$0.30 per question. AgentMemory injects ~2K relevant tokens — 90% savings.
It also solves "context pollution" — when there's too much info in the window,
the model gets confused and starts hallucinating. Less noise = more accurate answers.
Integrations
AgentMemory works with any MCP-compatible agent:
Claude Code, Cursor, Gemini CLI, Codex CLI, OpenClaw, OpenCode, Cline, Goose, Windsurf, and more.
All agents share the same memory server — what Claude Code learns, Cursor already knows.
Installation pitfalls — what NOT to do
openclaw mcp set only! — Manually editing openclaw.json with mcpServers or mcp.servers triggers schema validation failure. OpenClaw wipes the config every restart (saves as .clobbered)- No plugins integration — The
integrations/openclaw folder in the repo requires plugins.slots and plugins.entries which trigger the same clobber. MCP-only approach works perfectly
- ClawHub ≠ AgentMemory —
clawhub install agentmemory installs a different product (agentmemory.cloud, paid cloud service). The right repo: rohitg00/agentmemory — self-hosted and free
- iii-engine silent crashes — systemd reports "active" but the health endpoint is dead. Fix:
KillMode=mixed in the service to prevent orphan processes
Full installation prompt here — 5 steps, ready to copy.
Original project: rohitg00/agentmemory.
Bottom line: MCP is the most important thing to happen to AI agents since the context window.
Instead of throwing away knowledge at the end of every session, your agent actually remembers. And that changes everything.
Just read the README first 😉