Persistent · Semantic · Shared · Governed

Agent memory thatsurvives the session.

Context windows reset. MemClaw gives your AI agents a durable memory layer — write once, recall by meaning, share across the fleet— with rules and provenance built in.

Get Started Free →

No credit card

Open-Source Repo

Apache 2.0 · 70,000 downloads

In production at eToro — read the case study →

memory that outlives the session
# Tuesday — the deploy agent learns something
memclaw_write("Postgres 16 upgrade blocked by
  pgvector 0.6 — staying on 15 until Q3")

# six weeks later — a DIFFERENT agent asks
memclaw_recall("why are we still on postgres 15?")

→ "Postgres 16 upgrade blocked by pgvector 0.6…"
   fact · written by deploy-bot · 41 days ago
One memory layer — every agent, every session

Your agents are smart. Their memory isn’t.

Three reasons “just use the context window” stops working the day you have real agents doing real work.

The context window is not memory

Everything your agent figured out this session — decisions, dead ends, tribal knowledge — evaporates when the session ends. Tomorrow it re-derives all of it, at your token cost.

RAG is not memory either

Document retrieval answers "what do the docs say?" — not "what did we decide, what did we try, and how did it turn out?". Memory needs writes, updates, and a lifecycle, not just search.

N agents = N amnesiacs

Each agent in your fleet re-learns what a teammate already paid to learn. Without a shared layer, knowledge is trapped per-agent, per-session, per-tool.

Agent memory, done properly — not a vector database with homework

MemClaw ships the whole memory layer: storage, enrichment, recall, sharing, and the governance your team will eventually wish it had from day one.

Persistent & typed

Write once — MemClaw enriches every memory with type (fact, decision, episode…), title, summary, and tags automatically.

Semantic recall

Ask in plain language, get the right memory back — by meaning, not keywords, with temporal hints ("what changed last week?").

Shared across the fleet

Agent, fleet, and tenant scopes decide who sees what. What one agent learns, every agent that should know — knows.

Governed, with provenance

Keystone policy rules every agent obeys, trust levels that gate destructive ops, supersede-don't-delete corrections, and a full audit trail of who wrote what.

Plugs into anything

MCP-native: paste one config and 12 memory tools appear in Claude Code, Claude Desktop, Cursor, Windsurf. No MCP client? Plain REST works too.

Open source, or managed

Apache 2.0 — self-host in ~5 minutes. Or use MemClaw Cloud and deploy nothing. Same API either way.

Get Started Free →

Free tier · no credit card

Long-term memory for LLMs — any model, any client

LLM memory that isn’t welded to one vendor: the store lives outside the model, so it survives session resets, model swaps, and provider changes.

  • Model-agnostic by design

    Claude, GPT, Gemini, or local — if it speaks MCP or HTTP, it remembers. Swap models without losing a single memory.

  • Stop replaying transcripts

    Long-term LLM memory means recalling the three facts that matter instead of re-feeding yesterday's whole conversation at token cost.

  • Survives everything

    Session ends, context compacts, model upgrades — the memory layer doesn't notice. Write once, recall months later.

MCP—Any Client
{
  "mcpServers": {
    "memclaw": {
      "url": "https://memclaw.net/mcp",
      "headers": {
        "X-API-Key": "mc_your_key"
      }
    }
  }
}
Claude Desktop, Claude Code, Cursor, Windsurf—paste config and go
Get Started Free →

Free tier · no credit card

A memory architecture with a knowledge graph inside

Three planes, one API: a typed store, an entity graph that connects what your agents know, and a governance layer that decides who may do what.

Typed memory store

Facts, decisions, episodes, actions — every write auto-enriched with title, summary, and tags, deduplicated, and superseded (never silently lost) when facts change.

Entity knowledge graph

Memories link the people, projects, and systems they mention. Ask for an entity and get its whole graph view; tune traversal depth per search profile.

Governance plane

Agent, fleet, and tenant scopes; trust levels that gate destructive operations; keystone policy rules every agent loads; full who-wrote-what provenance.

Want the internals? Read the full architecture — the memory pipeline, step by step.

Get Started Free →

Free tier · no credit card

Remembering in 30 seconds

Three steps. The longest one is typing your email.

1~30 seconds

Create a free account

Email or SSO. No credit card, no sales call.

2one click

Copy your API key

A tenant-scoped key (starts with mc_) is waiting on your dashboard.

3done

Connect your agents

One MCP config or one REST call — your agents start remembering.

Works where your agents work

Comparing options? MemClaw vs Mem0, Zep, and Letta — honest tables.

Frequently asked questions

What is agent memory?+

Agent memory is a persistent store an AI agent writes to and recalls from across sessions — facts it learned, decisions that were made, what was tried and how it turned out. Unlike the context window (which resets) or RAG (which only reads documents), an agent memory layer supports writes, semantic recall, updates as facts change, and sharing between agents.

How is agent memory different from RAG?+

RAG retrieves passages from documents you already have. Agent memory captures what happens as your agents work: decisions, outcomes, preferences, tribal knowledge that exists in no document. MemClaw gives memories a lifecycle — they're written, enriched, recalled semantically, superseded when facts change, and audited — which document retrieval can't do.

How is MemClaw different from a vector database?+

A vector database is a building block — you'd still have to build enrichment, memory types, dedup, scopes, sharing, policy, provenance, and an agent-facing API on top. MemClaw is the finished memory layer: agents get 12 ready tools over MCP (or REST), and teams get governance — scopes, trust levels, keystone rules, audit — out of the box.

Can multiple agents share one memory?+

Yes — that's the point. Memories carry scopes (agent, fleet, tenant), so a finding written by one agent is instantly recallable by every agent that should see it, with the writer's identity attached. Per-agent API keys keep provenance real: you always know which agent learned what, and when.

How do I add memory to my agent?+

If your agent runs in an MCP client (Claude Code, Claude Desktop, Cursor, Windsurf), paste one config block and the memory tools appear. For frameworks like LangChain, LlamaIndex, CrewAI, or AutoGen there are integration guides, and the REST API works from any language. First write to first recall is typically under five minutes.

Is MemClaw open source? What does it cost?+

The engine is Apache 2.0 open source — self-host it in about five minutes. MemClaw Cloud has a free-forever tier (10K memories, unlimited agents and fleets, no credit card); Pro starts at $49/mo. Same API on both, so you can start managed and move self-hosted, or the reverse.

Give your agents a memory today.

Free tier: 10K memories, unlimited agents and fleets. No credit card.

Get Started Free →

No credit card

Open-Source Repo

Apache 2.0 · 70,000 downloads