Skip to content
Naveen Raj edited this page Apr 11, 2026 · 1 revision

Memory

Synapse agents have persistent memory — they can store and retrieve facts, relationships, and context across sessions and runs.


How It Works

Memory is implemented as a knowledge graph via the built-in @modelcontextprotocol/server-memory MCP server. This server starts automatically with Synapse — no configuration required.

Agents can:

  • Store facts, relationships, and entities
  • Retrieve relevant memories in future sessions
  • Update existing knowledge
  • Delete outdated information

The knowledge graph persists in the data directory and survives restarts.


Automatic Activation

The Memory MCP server is enabled by default. Every agent has access to memory tools unless you explicitly remove them from the agent's tool list.

Memory tools available to agents:

  • memory__create_entities — store new facts or entities
  • memory__create_relations — link entities together
  • memory__search_nodes — retrieve relevant memories by query
  • memory__open_nodes — fetch specific entities
  • memory__delete_entities — remove outdated information

What Agents Remember

Agents can store any structured knowledge:

People and organizations:

Entity: "Acme Corp"
Type: Company
Properties: industry=SaaS, size=50 employees, primary_contact=John Smith

Project state:

Entity: "Project Alpha"
Relations: assigned_to → "Research Agent", status → "in_progress"

Facts and context:

Entity: "API Rate Limits"
Properties: endpoint="/api/data", limit=100 requests/min, last_checked=2024-01-15

Practical Use Cases

Cross-session continuity: An agent researching a topic can remember what it already found, avoiding redundant work in future sessions.

User preferences: A personal assistant agent remembers your preferences (preferred email format, timezone, recurring tasks) without you having to re-state them.

Accumulated knowledge: A data analyst agent remembers which database tables are relevant to your business, what queries performed well, and what anomalies it found previously.

Project tracking: An engineering agent tracks which issues it has investigated, what solutions were attempted, and what's still outstanding.


Sequential Thinking

The Sequential Thinking MCP server (also enabled by default) complements memory by giving agents structured reasoning:

  • Agents break complex tasks into explicit thought chains before acting
  • Each step is reasoned through before tool calls are made
  • Helps with multi-step problems that require planning

This is separate from memory — Sequential Thinking helps with within-session reasoning, while Memory persists knowledge across sessions.


Viewing Memory

Memory is stored in the data directory. You can inspect and manage it via:

  • Settings → Memory (if a memory management UI is available in your version)
  • Directly querying the knowledge graph through an agent: "What do you remember about [topic]?"

Disabling Memory

To prevent an agent from using memory tools, edit the agent in Settings → Agents and remove the memory__* tools from its tool list.

To disable the Memory MCP server entirely, remove it from Settings → MCP Servers.


See Also

  • Agents — configuring agent tool access
  • Built-in Tools — all built-in MCP servers
  • Vault — file-based persistence (complements memory)

Clone this wiki locally