AsanagiAMI

04. AI Memory

Don't just know what happened. Know why.

Your model has amnesia.
Until now.

Every conversation starts fresh. Every context window expires. Every insight you built together evaporates when the session ends. You're not collaborating with an AI — you're briefing a new contractor every morning.

AsanagiDB-MCP gives your model something it has never had before: long-term memory — a persistent knowledge graph that survives context resets, spans sessions, and keeps relationships between facts, not just the facts themselves.

And short-term memory — when a new conversation starts, your model is already warmed up. It knows what you were working on, what decisions were made, and what's still open. No briefing. No recap. The conversation picks up exactly where it left off.

No markdown files. No flat JSON dumps. A graph, where every memory has edges to related memories.

// Without AsanagiDB-MCP

User: "Hey, where were we?"

Model: "I don't have context from previous sessions. Could you catch me up?"

// With AsanagiDB-MCP — new session, zero briefing

User: "Hey, where were we?"

Model: "You were finalizing the logo mockups. Design approval is still pending — that's the blocker on the white paper too. Want me to draft a follow-up to the client?"

GraphRAG, No Extraction Pipeline.

No corpus. No extraction.
No lossy pipeline.

GraphRAG usually means retrofitting a graph onto text that already exists — an extraction pipeline reads documents, infers entities and relationships, and reconstructs structure the source never had explicitly. That's a workaround for data that started as prose. When the agent writing the memory has direct graph access, the workaround has nothing to fix: the relationship becomes an edge the moment it's known, not a sentence a pipeline parses back out later. Same retrieval-augmented pattern — query the graph, ground the answer — without the step that loses fidelity.

Typical GraphRAG

Text corpus has to exist first.
An extraction pipeline infers entities and relationships — lossy by construction.
The graph is a retrieval index, rebuilt periodically as the corpus changes.
The model is invoked at fixed pipeline steps, not writing the graph itself.

AsanagiDB

The agent has direct graph access.
Relationships are edges, authored at write time — no extraction step to lose fidelity to.
The graph is the memory, not an index over something else.
The model reads and writes the graph directly, every session.

The memory vocabulary.

memory_save

Write a node with body, type, and edges to related concepts.

memory_search

Fuzzy-match across all node names. Fast enough for real-time use.

memory_context

Load a node and its neighborhood — the full relational picture.

memory_relate

Draw a typed edge between two existing nodes.

memory_hottest

Surface the nodes accessed most recently and frequently. Session continuity.

memory_save_batch

Write multiple nodes in one round trip. Efficient for bulk context saves.

Connect in three lines.

// claude_desktop_config.json

{
  "mcpServers": {
    "AsanagiDB": {
      "command": "/usr/local/bin/AsanagiDB-mcp",
      "args": ["mcp"],
      "env": {
        "AsanagiDB_NAME": "memory"
      }
    }
  }
}

Works with Claude, Gemini, and any agent that speaks the Model Context Protocol. Same config format across all clients.

04. AI Memory

Your model has amnesia. Until now.

No corpus. No extraction. No lossy pipeline.

The memory vocabulary.

Connect in three lines.

Connected memory, with a price you can forget.

Your model has amnesia.
Until now.

No corpus. No extraction.
No lossy pipeline.