Studiomeyer Memory

Part of the StudioMeyer MCP Stack — Built in Mallorca 🌴 · ⭐ if you use it

StudioMeyer Memory

Persistent, intelligent memory for AI agents. 56 MCP tools (incl. interactive 3D knowledge graph visualization). Knowledge Graph included in every plan.

A note from us

We have been building tools and systems for ourselves for the past two years. The fact that this repo is small and has few stars is not because it is new. It is because we only just decided to share what we have built. It is not a fresh experiment, it is a long story with a recent commit.

We love building things and sharing them. We do not love social media tactics, growth hacks, or chasing stars and followers. So this repo is small. The code is real, it gets used, issues get answered. Judge for yourself.

If it helps you, sharing, testing, and feedback help us. If it could be better, an issue is more useful. If you build something with it, tell us at [email protected]. That genuinely makes our day.

From a small studio in Palma de Mallorca.

What is this?

StudioMeyer Memory gives AI agents persistent memory across sessions. Instead of starting fresh every conversation, your agents learn, remember, and improve over time.

56 MCP tools for learning, search, knowledge graph, session tracking, multi-agent support, contradiction detection, self-improvement — plus interactive 3D visualization (nex_graph_view, nex_recall_timeline, nex_session_replay) via MCP Apps.

Connect in 10 Seconds

Claude Desktop / Cowork

Settings → Connectors → Add URL:

https://memory.studiomeyer.io/mcp

Claude Code

claude mcp add --transport http memory https://memory.studiomeyer.io/mcp

Cursor / VS Code / Windsurf / Zed

npx mcp-remote https://memory.studiomeyer.io/mcp

REST API

For custom agents, scripts, and integrations:

# 1. Sign up
curl -X POST https://memory.studiomeyer.io/api/signup \
  -H "Content-Type: application/json" \
  -d '{"email": "[email protected]"}'

# 2. Store a memory
curl -X POST https://memory.studiomeyer.io/api/learn \
  -H "Authorization: Bearer nex_your_key" \
  -H "Content-Type: application/json" \
  -d '{"category": "insight", "content": "Users prefer dark mode"}'

# 3. Search
curl -X POST https://memory.studiomeyer.io/api/search \
  -H "Authorization: Bearer nex_your_key" \
  -H "Content-Type: application/json" \
  -d '{"query": "user preferences"}'

API Endpoints

Method	Endpoint	Description
POST	/api/signup	Create account, get API key
GET	/api/account	Account info, plan, limits
POST	/api/learn	Store a learning
POST	/api/search	Search all knowledge
POST	/api/decide	Log a decision
POST	/api/entity	Create/update entity
POST	/api/import	Import from ChatGPT/Claude/Gemini/Copilot/Perplexity
POST	/api/backfill	Generate embeddings
GET	/api/export	Export all data (Markdown)
POST	/api/checkout	Upgrade plan (Stripe)

Full API documentation: openapi.yaml

SDKs

• OpenAPI Spec: openapi.yaml — generate a typed client for any language from the spec.

Features

Core Memory

• 6-phase parallel search — semantic vectors (text-embedding-3-large @ 1024 dims since July 2026), trigram matching, full-text (DE/EN), graph traversal, temporal + contradiction search — fused via Reciprocal Rank Fusion, with optional cross-encoder reranking (Pro) and honest abstention ("I don't know" below the evidence threshold)
• Duplicate detection — 5-factor admission control prevents noise
• 10 learning categories — pattern, mistake, insight, research, architecture, and more
• Decision logging — track choices with reasoning for traceability

Knowledge Graph (All Plans)

• Entities — people, tools, projects, concepts
• Observations — facts attached to entities with temporal validity
• Relations — typed, directional connections between entities
• Graph traversal — 2-hop BFS search through connections
• Temporal queries — "What was true on March 15th?"

Intelligence

• Contradiction Detection — automatically flags conflicting facts
• Trust Weighting — verified memories win near-ties over unverified agent claims (memory-poisoning defense, July 2026)
• Multi-tenant-safe vector search — hnsw.iterative_scan keeps tenant/project filters from silently swallowing results; an embedding-dimension guard makes config drift loud instead of silently degrading search
• FadeMem Adaptive Decay — important memories fade 6x slower
• Confidence Scoring — 0-1 scores with temporal decay
• Bi-temporal Model — event time + ingestion time
• Auto-Consolidation — merges duplicates, backfills embeddings

Multi-Agent

• Agent namespaces — scope memories to specific agents
• Cross-agent search — search across all agents by default
• Shared Knowledge Graph — one graph, multiple agents

Import

• 5 platforms — ChatGPT, Claude, Gemini, GitHub Copilot, Perplexity
• LLM extraction — entities, learnings, decisions, relations
• Preview mode — analyze before importing

Benchmark — LongMemEval

86% on LongMemEval 50q stratified, GPT-4o judged (Run S957b, 1 May 2026, Anthropic Sonnet 4.6 as answer generator).

Per-category breakdown:

Category	Score
temporal-reasoning	92%
multi-session	92%
single-session-user	100%

Caveat for honesty: S957b ran against memory-server v3.16.10, before the v3.16.11 cross-project search-leak fix landed (LIVE on production since 2 May 2026). The 86% may be 1-3 percentage points inflated by cross-question help. A clean 50q re-run with the fix in place is the next planned wave once cost-control safeguards in the bench script are wired up. Realistic range after the re-run: 78-86%.

The full whitepaper with methodology, run history and per-question results lives at studiomeyer.io/services/memory and is updated as new validated runs land.

Reference scores from other vendors (publicly published):

System	Aggregate	Temporal	Source
StudioMeyer Memory	86%	92%	this README
Mem0 (Managed Platform)	93.4%	93.2%	mem0.ai/blog April 2026 — Token-Efficient Memory Algorithm
Mem0 BEAM 10M tokens	48.6%	16.3%	mem0.ai/research — production-scale collapse
Mastra Observational Memory	94.87% (gpt-5-mini)	—	mastra.ai/blog/observational-memory
Hindsight	91.4%	79.7%	published
Zep	63.8%	—	arXiv 2501.13956

What we are honest about: in aggregate we sit below Mem0 Managed (93.4%) and Mastra OM with gpt-5-mini (94.87%). Our differentiators are architecture and production behavior — bi-temporal knowledge graph included from the Free tier (Mem0 charges $249/mo for graph features), production-scale stability (Mem0 itself reports 48.6% aggregate / 16.3% temporal at 10M tokens), EU Nuremberg storage + GDPR, and the live interactive 3D knowledge graph as part of the standard tool surface — verified unique among AI memory SaaS as of April 2026.

Pricing

Plan	Price	Calls/Day	Learnings	Entities	Knowledge Graph	Agentic Search + AI Reranker
Free	€0 forever	5,000	1,000	100	✅	Simple hybrid search
Pro	€9/mo	50,000	25,000	1,000	✅	✅
Team	€19/mo	Unlimited	Unlimited	Unlimited	✅	✅

All 56 tools available in every plan — including interactive 3D knowledge graph visualization. Knowledge Graph included from Free.

What's the difference? Free uses our hybrid search stack (Trigram + FTS + pgvector + Knowledge-Graph traversal + RRF fusion + temporal decay + per-tenant isolation) — fast, deterministic, no LLM cost. Pro+ adds agentic retrieval: an iterative completeness check (Haiku LLM) that detects when a complex query (multi-hop, temporal computation, aggregation, list ordering) needs a refined re-query. On LongMemEval-class benchmarks this lifts recall by 5-10pp on the hardest categories.

Compare: Mem0 charges $249/mo for Graph Memory. We include it from Free.

Sessions, reads and corrections

Pass a stable conversationId with every MCP call in one conversation. Different conversations under the same account use separate sessions. agentId attributes or filters knowledge; it is not a conversation boundary. Clients without conversationId retain the legacy session behavior.

Use readOnly: true on read operations to avoid memory and session mutations. trackUsage: false on retrieval also disables session bookkeeping and memory reinforcement. SaaS request quotas and security accounting still apply. Write operations with readOnly: true return an error.

Similarity alone does not authorize replacing a fact. Non-identical near matches are preserved as separate learnings with update guidance. To correct a fact, call nex_learn_update with its exact id and replacement content; short corrections are valid. The response reports whether the new embedding was stored or is pending. Search rank is a ranking score, not a percentage; abstention reports its reason separately from the returned candidates.

To permanently delete one learning, use nex_learn_archive with action: "delete", the exact id and confirm: true. Archiving remains reversible; deletion does not.

Product facts and release checks

product-manifest.json is generated from the same source that enforces server limits and registers MCP tools. The REST collection documents REST operations; it is not a collection of one request per MCP tool.

Run node scripts/sync-product-facts.mjs --check before publishing documentation. On a new server release, export its manifest, run the script without --check, and review the resulting diff. Daily plan limits are separate from the configurable MCP request burst limit (default 60 requests per minute per tenant).

Security

• Magic Link Authentication — email verification on every sign-in. No passwords stored. You receive a single-use link that expires in 10 minutes. Nobody can access your data without proving email ownership.
• OAuth 2.1 with PKCE S256 — latest authentication standard
• Own Postgres in the EU (Hetzner Nuremberg, Germany) — GDPR-ready, no third-party database provider
• Tenant-scoped access: application queries bind data to the authenticated tenant; database RLS is not currently the enforcement layer.
• Secret Scanner — prevents accidental storage of API keys/passwords
• Data Export — GET /api/export anytime
• Transport encryption: HTTPS for the hosted endpoint.
• AI processing: embeddings and optional model features send relevant content to configured providers. EU database storage does not imply EU-only model processing. Secret scanning is not general PII anonymization.

Examples

See the examples/ folder:

• Python basic usage
• LangChain integration
• CrewAI multi-agent
• TypeScript basic
• ChatGPT import
• Official SDK usage

Resources

• Documentation: Built-in — call nex_guide("quickstart") via MCP
• Postman Collection: Download
• OpenAPI Spec: openapi.yaml
• MCP Registry: io.studiomeyer/memory

About StudioMeyer

StudioMeyer is an AI and design studio based in Palma de Mallorca, working with clients worldwide. We build custom websites and AI infrastructure for small and medium businesses. Production stack on Claude Agent SDK, MCP and n8n, with Sentry, Langfuse and LangGraph for observability and an in-house guard layer.

License

MIT — see LICENSE. The memory server is a hosted service — this repository contains documentation and examples.

Links

• 🌐 memory.studiomeyer.io
• 🏢 studiomeyer.io