Elefant — Persistent memory for your AI

Every chat starts cold.
Every context gets re-typed.

Without a memory layer, you re-explain your stack on Monday, re-decide naming conventions on Wednesday, and lose every hard-won troubleshooting fix the second the window closes. Elefant is the layer that makes Claude behave as if it remembers — across web, desktop, mobile, Claude Code, and any agent you wire up.

What you get

A persistent brain, on your own infrastructure.

Designed for the agentic workflow. Tuned for high-signal recall. Hosted by you.

Semantic search, two stores

Embedding-indexed memories and structured wiki pages, searchable together in one ranked call. Filter by tags. Pivot between atomic facts and long-form references without leaving the protocol.

MCP-native by design

Speaks Model Context Protocol over HTTP with Bearer auth. Plugs into Claude Desktop, Claude Code, Claude in Chrome, and any client that supports MCP — no glue code, no proprietary SDKs.

Hosted today, self-run soon

Start in minutes on managed Elefant SaaS — we run the stack, you keep your tokens. Or wait for the containerized release and run the entire system inside your own network. Same protocol either way.

Built for write discipline

A typed memory model — fact, decision, preference, procedure, event — plus a controlled tag vocabulary. Update over duplicate. The store gets sharper over time, not noisier.

One memory, every surface

The Claude on your phone and the Claude in your terminal pull from the same source of truth. Decide something on the train; your CLI agent knows about it before you reach the desk.

Wiki, for when memories aren't enough

Long-form pages for protocols, project overviews, and reference material — versioned, image-aware, and reindexable on demand. Atomic recall and structured reading, in one platform.

The mental model

Two stores. One source of truth.

Atomic facts retrieved on demand. Structured knowledge read top-to-bottom. Both indexed, both searchable, both yours.

🧠 memories

Discrete, atomic, tagged.

Decisions, preferences, outcomes, procedures. The things you want to retrieve in a single line — not read as a document.

Six types: fact, decision, preference, procedure, event, derived
Embedding-indexed, tag-filterable
Updateable — supersede stale entries in place
Self-contained, readable in isolation

📖 wiki

Long-form, structured, image-aware.

Protocols, project overviews, reference material. The things you'd want to read end-to-end and link between.

Markdown pages with frontmatter
Embedded images fetched inline on retrieval
Path-based addressing, deliberate editing
On-demand vector reindex after bulk changes

Operating principles

High-signal, on purpose.

Elefant's protocol is opinionated, because the failure mode of memory layers is becoming a junk drawer of half-relevant entries.

Update over duplicate

Before adding, search. Refining or contradicting an existing memory? Update it in place — don't fragment the signal across three overlapping entries.

Controlled vocabulary

Tags are how memories get grouped. A defined set — project, domain, type — keeps the corpus coherent instead of bloating into one-off labels.

Write the durable

Decisions, preferences, outcomes, procedures, identifiers, constraints. Not chat transcripts, not transient questions, not secrets. Distillation, not logging.

Read before answering

When the question touches your work, search Elefant first. A cheap lookup beats a confidently wrong answer that ignores three months of context.

Two ways to run Elefant

Hosted today. Self-run when you're ready.

Same protocol, same data model, same MCP surface. Pick the deployment that matches where your team is right now — and migrate later without rewriting a single tool call.

Available now

Elefant SaaS

Managed by Zennthic. Provisioned in minutes. Production-ready from day one.

Zero-ops onboarding. Sign up, get an MCP endpoint and a Bearer token, point Claude at it.
Managed embeddings & vector store. No model files to download, no infrastructure to size, no upgrades to run.
Per-workspace isolation. Token-scoped labels keep projects, teams, and clients separate.
Continuous updates. New tools, retrieval improvements, and protocol features ship to your instance without a maintenance window.
Export when you're ready to move. Same data model as self-hosted — your corpus comes with you.

Start free trial

Shipping soon

Elefant Self-Hosted

A single containerized stack you run inside your own network, on your own keys.

One Docker stack. API, embeddings, vector store, and wiki — orchestrated as a single deployable unit.
Your data never leaves. Memories, wiki pages, embeddings, and tokens stay on your hardware. End of story.
Offline-validated licensing. ES256 JWT verified at startup — no license server contacted, no phone-home telemetry.
Tiered builds, one image. Switch between Free CE and Premium with an environment variable — no separate distributions to manage.
Air-gap friendly. Designed for regulated environments. Bring your own IdP, bring your own egress policy.

Get notified at launch

Every Claude. One brain.

Connect once. Recall everywhere.

Elefant runs as an MCP server over HTTP. Any client that speaks MCP can read and write — under whatever token scope you grant.

Claude.ai — Web, Desktop & Mobile

Connector-based MCP. Full read/write per chat or project.

Claude Code

Configured via ~/.claude/settings.json with HTTP transport + Bearer auth.

Claude in Chrome

Browser-side recall while you research and ship.

Custom & API-based agents

Anything that speaks MCP. Token-scoped, label-filtered.

~/.claude/settings.json

// Connect Claude Code to your Elefant instance
{
  "mcpServers": {
    "elefant": {
      "transport": "http",
      "url": "https://elefant.your-domain/mcp",
      "headers": {
        "Authorization": "Bearer ${ELEFANT_TOKEN}"
      }
    }
  }
}

# Or one-shot:
$ claude mcp add --transport http elefant \
    https://elefant.your-domain/mcp \
    --header "Authorization: Bearer $ELEFANT_TOKEN"

Your AI forgets everything.
Elefant doesn't.