Synaptic Graph

A portable, human-memory-inspired memory layer for AI systems. Your AI conversations generate knowledge — Synaptic Graph captures it, connects it, and makes it available everywhere.

What it does

Synaptic Graph is an MCP server that gives any AI assistant persistent, portable memory. Memories are stored as a weighted graph of impulses (learned things) connected by relationships that strengthen with use and fade with neglect — just like human memory.

Key features

• Spreading activation retrieval — queries light up matching memories and propagate through connections, surfacing related context you didn't explicitly search for
• Weighted decay — memories strengthen when accessed and fade when neglected. Nothing is deleted — even old memories can resurface with the right trigger
• Ghost graphs — overlay your existing knowledge bases (Obsidian vaults, markdown directories) without copying or modifying them
• Tags and source tracking — organize memories by topic, see which AI provider contributed each memory
• Auto-linking — new memories automatically connect to existing ones via keyword overlap
• Obsidian export — export your memory graph as linked markdown files with frontmatter and wikilinks
• Desktop app — Tauri-based galaxy visualization with interactive force-directed graph
• Incognito mode — full blackout, zero trace when you need privacy

MCP Tools (28)

Memory: save_memory, quick_save, retrieve_context, recall_narrative, update_memory, delete_memory, inspect_memory, confirm_proposal, dismiss_proposal, list_candidates, propose_memories, explain_recall

Graph: link_memories, unlink_memories, register_ghost_graph, refresh_ghost_graph, pull_through, export_to_obsidian

Tags: create_tag, list_tags, tag_memory, untag_memory

System: memory_status, set_incognito, create_backup, sync_export, sync_status

Install

Quick install

git clone https://github.com/jwgrogan/synaptic-graph.git
cd synaptic-graph
./install.sh

The installer builds from source and configures your AI clients automatically.

Manual setup

Build

cargo build --release

Claude Code

Add to ~/.mcp.json:

{
  "mcpServers": {
    "synaptic-graph": {
      "command": "/path/to/synaptic-graph/target/release/synaptic-graph"
    }
  }
}

Claude Desktop

Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "synaptic-graph": {
      "command": "/path/to/synaptic-graph/target/release/synaptic-graph"
    }
  }
}

Codex

Add to ~/.codex/config.json:

{
  "mcpServers": {
    "synaptic-graph": {
      "command": "/path/to/synaptic-graph/target/release/synaptic-graph"
    }
  }
}

The AGENTS.md file in this repo provides Codex with instructions for how to use the memory tools.

Gemini CLI

Add to your Gemini CLI MCP config. The GEMINI.md file in this repo provides Gemini with instructions for how to use the memory tools.

OpenClaw

OpenClaw uses the same MCP protocol as Claude Code. Add to your OpenClaw MCP config:

{
  "mcpServers": {
    "synaptic-graph": {
      "command": "/path/to/synaptic-graph/target/release/synaptic-graph"
    }
  }
}

Any MCP-compatible client

synaptic-graph is a standard MCP server over stdio. Point any MCP client at the binary — no additional configuration needed. The server embeds its own instructions via the MCP ServerInfo.instructions field.

Desktop app

cd ui
npm install
npx tauri dev     # development
npx tauri build   # production .dmg/.exe

CLI

The binary also works as a CLI:

synaptic-graph status              # show memory stats
synaptic-graph retrieve "topic"    # search memories
synaptic-graph save "content"      # save a memory

How it works

Memories are stored as impulses — atomic units of knowledge (not raw conversation transcripts). When you save a memory, the system:

1. Redacts secrets and PII
2. Creates a candidate impulse
3. Auto-links to existing memories via keyword overlap
4. Confirms and indexes for full-text search

When you retrieve context, spreading activation propagates through the weighted graph — directly matching memories activate first, then energy spreads through connections to surface related context. Frequently accessed connections strengthen. Unused ones fade.

Architecture

• Rust core with SQLite (single portable file)
• MCP server over stdio (works with any MCP-compatible client)
• Tauri v2 desktop app with Svelte + Pixi.js
• 148 tests across all modules

Documentation

• Philosophy — the human-memory-inspired design thinking
• PRD — product requirements and roadmap
• TRD — technical architecture and data model

License

Apache 2.0 — see LICENSE