Synix

Synix Plugin for Claude Code

Persistent cross-session memory for Claude Code via a Synix knowledge server.

What It Does

• Session start hook — fetches the synthesized context document from the Synix server and injects it into the conversation, so Claude starts every session with full memory
• Session end hook — pushes the session transcript (.jsonl) to the Synix server for ingestion into the next build cycle
• MCP server — connects to the Synix knowledge server's MCP endpoint, exposing search, ingest, context retrieval, and bucket listing tools
• Memory skill — teaches Claude when and how to use the memory tools (search prior work, ingest notes, retrieve context)

Prerequisites

• A running Synix knowledge server (see Synix docs)
• Network access to the server from your machine (e.g. same LAN, Tailscale, SSH tunnel)
• bash, curl, python3 (standard on macOS and Linux)
• Claude Code CLI installed (npm install -g @anthropic-ai/claude-code)

Installation

Quick start (development / testing)

Run Claude Code with the --plugin-dir flag pointing to this directory:

claude --plugin-dir /path/to/synix/plugin

Persistent install

1. Add the synix repo as a plugin marketplace (one-time):

   In Claude Code, run:

   /plugin marketplace add https://github.com/marklubin/synix
   

2. Install the plugin:

   /plugin install synix@synix --scope user
   

   Scope options:

   • user — available in all projects on this machine
   • project — available only in the current project (stored in .claude/settings.json)
   • local — available only for you in this project (stored in .claude/settings.local.json)

3. Reload plugins in your current session:

   /reload-plugins
   

Verify installation

Run /plugin in Claude Code and check the Installed tab — you should see "synix" listed and enabled.

Then test the MCP connection:

Ask Claude: "Can you list the available memory buckets?"

If the server is reachable, Claude will call list_buckets() and return the configured buckets.

Configuration

During installation, the plugin prompts for your Synix knowledge server URL. This is stored in your Claude Code settings and used by both hooks and the MCP server.

The default server URL is http://salinas:8200.

You can also override the server URL with the SYNIX_SERVER_URL environment variable:

SYNIX_SERVER_URL="http://localhost:8200" claude

The resolution order is: CLAUDE_PLUGIN_OPTION_SERVER_URL (plugin config) → SYNIX_SERVER_URL (env var) → http://salinas:8200 (default).

How It Works

Session start (hooks/session_start.sh)

On every new session, the hook calls GET /api/v1/flat-file/context-doc to fetch the current synthesized context document. This is injected into the conversation as additionalContext, so Claude immediately has full memory context — recent activity, open threads, research state, user model, and work status.

If the server is unreachable, the hook silently succeeds with no context injected. Sessions are never blocked by a down server.

Session end (hooks/session_end.sh)

When a session ends, the hook:

1. Reads the sessionId from the hook input JSON
2. Locates the session transcript at ~/.claude/projects/*/{sessionId}.jsonl
3. POSTs the transcript content to POST /api/v1/ingest/sessions

This feeds the session into the Synix build pipeline. On the next auto-build cycle, the server will process it through episode summarization, fold synthesis, and context generation.

If the server is unreachable or the session file can't be found, the hook silently succeeds.

MCP tools

The MCP server exposes the full Synix tool surface. Key tools for memory:

Tool	Purpose
search(query, layers?, limit?)	Hybrid search across memory (fulltext + semantic)
ingest(bucket, content, filename)	Write content to a source bucket
get_context(name?)	Retrieve a synthesized context document
list_buckets()	List available ingestion buckets

Memory skill (skills/memory/SKILL.md)

The skill teaches Claude when to proactively use memory tools:

• Search when the user references prior work, decisions, or context
• Ingest after calls, meetings, or when the user says "remember this"
• Get context when starting a complex task where historical context helps

Structure

plugin/
  .claude-plugin/
    plugin.json              Plugin manifest (name, version, pointers to hooks/mcp)
  .mcp.json                  MCP server configuration (streamable-http to synix)
  hooks/
    hooks.json               Hook registration (SessionStart, Stop)
    session_start.sh         Fetches context on session start
    session_end.sh           Pushes transcript on session end
  skills/
    memory/
      SKILL.md               Memory skill definition
  README.md                  This file

Troubleshooting

"Cannot connect to MCP server"

• Verify the Synix server is running: curl -sf http://your-server:8200/api/v1/buckets
• Check network access (Tailscale, firewall, etc.)
• Verify the URL in .mcp.json matches your server

Sessions not appearing in memory

• Check that the session end hook is registered: run /plugin and look for hook errors
• Verify the server is reachable from the hook: curl -sf http://your-server:8200/api/v1/buckets
• Check server logs for ingest errors
• Auto-build has a cooldown (default 300s) — the session won't appear in search until the next build completes

No context injected at session start

• The context-doc projection must exist — run a build + release on the server first
• Check that SYNIX_SERVER_URL is set correctly (or that the default http://salinas:8200 resolves on your network)