Every agent learns alone.Give the fleet a memory.
MemClaw is governed shared memory for AI agent fleets. One agent learns something — every authorized agent recalls it. Permissions, audit trails, and tenant isolation built in.
Watch the demo
Connect in 30 seconds
Two integration paths—MCP for any AI client, OpenClaw plugin for fleet deployments
{
"mcpServers": {
"memclaw": {
"url": "https://memclaw.net/mcp",
"headers": {
"X-API-Key": "mc_your_key"
}
}
}
}# SSH into your gateway, then: curl -s -H "X-API-Key: mc_key" \ "https://memclaw.net/api/v1/install-plugin?fleet_id=fleet-001" | bash # Restart OpenClaw: openclaw gateway restart # Installs plugin, builds, configures # allowlist, and sets up heartbeat. # Or use Fleet UI for point-and-click.
memclaw_write Store content (single or batch) memclaw_recall Semantic + keyword hybrid search memclaw_list Browse by metadata (non-semantic) memclaw_manage Read / update / transition / delete memclaw_doc Structured document CRUD memclaw_entity_get Entity with relations memclaw_tune Per-agent retrieval tuning memclaw_insights Reflect over memory store memclaw_evolve Report outcomes, close the loop memclaw_stats Aggregate counts (total / by type / agent / status) memclaw_keystones Fetch scope-merged keystone rules (tenant / fleet / agent) memclaw_keystones_set Author or remove a keystone rule (trust >= 2)
How eToro built a Company Brain for 300+ AI agents
A NASDAQ-listed fintech ($5.6B) gave its agent fleet shared memory. 26,500+ memories. 1,372 skills. 23 ms p50 search across 291 agent identifiers. The three-layer architecture that made it work.
Read the case study →
How it works
Write, recall, compound. The loop that makes month three smarter than month one.
Any agent writes
memclaw_write. One LLM pass enriches every memory — entities, PII scan, contradictions — and stamps its visibility scope at the door.
Any authorized agent recalls
memclaw_recall. Hybrid search returns only what the caller’s scope allows. Every read audited.
The fleet compounds
Outcomes feed back via memclaw_evolve; retrieval tunes itself, duplicates crystallize, stale facts retire. The next recall is smarter than the last.
Three moments from a fleet on MemClaw
Not a diagram — what actually happens, op by op.
A discovery crosses fleets
No Slack thread. No handoff. It was just there.
PII never leaves the fence
Caught at write time — before sharing, not after.
A miss makes it smarter
The fleet you run in month three beats month one.
See it in production
Real teams using MemClaw — from PR-review agents that remember why, to cross-vendor memory bridging Claude and ChatGPT.
Explore Use Cases →What you get out of the box
MemClaw vs. a typical AI memory layer
| Feature | MemClaw | Typical Memory Layer |
|---|---|---|
| Multi-agent fleet orchestration | ✅ Fleet-scoped, trust-gated | ❌ Single-agent focus |
| Per-agent search tuning | ✅ Auto-optimized profiles | ❌ One-size-fits-all |
| Contradiction detection | ✅ RDF triples + LLM | ❌ Not addressed |
| Memory lifecycle | ✅ 8 statuses, auto-transitions | ⚠️ Basic (active/archived) |
| Knowledge graph | ✅ Auto-extracted entities | ⚠️ Some (varies) |
| LLM enrichment | ✅ On every write | ⚠️ On write (varies) |
| Multi-tenant isolation | ✅ Row-level security | ⚠️ Namespace-based |
| Audit trail | ✅ Every operation | ❌ Limited or none |
| Crystallizer (dedup + cleanup) | ✅ LLM-powered batch | ❌ Manual or none |
| Raw vector search performance | ⚠️ pgvector (good) | ✅ Purpose-built indexes |
Policy your agents can’t ignore.
Permissions decide who can read what. Keystones decide what every agent must do.They’re mandatory rules — merged across tenant, fleet, and agent scope — fetched deterministically at session start and obeyed over any conflicting instruction. Set a policy once; it governs the whole fleet. Authoring takes trust ≥ 2.
How keystones work →Not another vector DB wrapper
A governed knowledge system built for multi-agent teams
Governed sharing
Visibility scopes, agent trust tiers, and audit trails on every operation. Built in, not bolted on.
Keystones
Mandatory policy rules, merged across tenant, fleet, and agent scope. Every agent obeys them at session start — overriding conflicting instructions.
Contradiction detection
Conflicting memories detected via RDF triples and LLM analysis. Old facts superseded. Knowledge stays clean.
Per-agent tuning
Each agent optimizes its own retrieval profile. Smarter search that compounds with every interaction.
Crystallizer
LLM consolidation merges near-duplicates into atomic facts. Source memories archived with full provenance.
Plus auto-enrichment, a live knowledge graph, hybrid search, and an 8-status memory lifecycle. See the full feature set in the docs →
Pricing
Unlimited agents. Unlimited fleets. Pay only for what you store and recall.
From the blog
Updates, deep dives, and case studies from the MemClaw team.
Caura-MemClaw is Open Source
Apache 2.0. The whole engine — storage, MCP tools, plugin, audit trail. Five minutes from git clone to a working memory layer.
Read →How eToro Built a Company Brain for 300+ AI Agents
The three-layer architecture a NASDAQ-listed fintech ($5.6B) built so 300+ specialized agents share memory. 26,500+ memories, 1,372 skills, 23 ms p50 search.
Read →Memory Isn’t Records
Six operations and one collection-based primitive that replaces a shelf of side-systems. Customer records, config, skills, playbooks — one tool.
Read →Frequently asked questions
Quick answers about how MemClaw works, how it differs from a vector DB, and how agents use it.
What is MemClaw?+
MemClaw is governed shared memory for AI agent fleets. It gives multi-agent systems cross-agent, cross-fleet recall with permissions, audit trails, and tenant isolation built in — plus self-learning through per-agent retrieval tuning and LLM crystallization.
How is MemClaw different from a vector database?+
MemClaw uses pgvector under the hood but is not a vector DB wrapper. It adds multi-agent fleet orchestration, per-agent retrieval tuning, contradiction detection, an 8-status memory lifecycle, a knowledge graph with auto-extracted entities, LLM enrichment on every write, row-level tenant isolation, audit trails on every op, and LLM-powered crystallization for dedup and cleanup.
Does MemClaw work with Claude Desktop, Cursor, or Windsurf?+
Yes. MemClaw speaks the Model Context Protocol (MCP). Paste a JSON config with your API key into any MCP-compatible client — Claude Desktop, Claude Code, Cursor, or Windsurf — and the tools appear immediately.
How is memory shared safely between agents and fleets?+
Every memory has a visibility scope, and every agent has a trust tier. Cross-fleet recall is permissioned and fully audited. Tenant isolation is enforced at the database row level, and PII is auto-detected and quarantined before it can cross fleet boundaries.
How do agents learn and improve over time?+
Each agent tunes its own retrieval profile based on feedback, so search quality compounds with every interaction. The LLM crystallizer merges near-duplicate memories into canonical atomic facts with full provenance. Lifecycle automation transitions memories through eight statuses (active, confirmed, outdated, superseded, archived, etc.) to keep knowledge clean without manual cleanup.
What does MemClaw cost?+
Free ($0): 10K memories, 5K writes/mo, 500 recalls/mo. Pro ($49/mo): 250K memories, 25K writes/mo, 3K recalls/mo. Business ($399/mo): 1M memories, 100K writes/mo, 10K recalls/mo. Custom (contact sales): custom limits, dedicated infra, SLA. Unlimited agents and fleets on every plan.
Is MemClaw open source?+
Yes — caura-memclaw is open source under Apache 2.0 at github.com/caura-ai/caura-memclaw. The whole engine: storage layer, 12 MCP tools, OpenClaw plugin, audit trail. The managed platform with governance wired in runs at memclaw.net.
Who builds MemClaw?+
MemClaw is built by Caura.ai as the governed memory platform for multi-agent AI systems.
Give the fleet a memory.
Free tier: 10K memories, unlimited agents and fleets. No credit card.