Memory
Synapse provides session memory (short-term, per-conversation) and two approaches for persistent memory across sessions: Vault (explicit, structured) and the native memory tool (model-managed, automatic).
Session memory
Every conversation is scoped to a session_id. Synapse stores the full message history for each session and replays it as context on every request.
- Session IDs are generated by the frontend on page load (or passed by the API caller)
- Each browser tab gets a fresh session by default — reload = new session
- The API caller can pass
"session_id": "my-session"to maintain continuity across requests
Context window limits
For long conversations, the message history can exceed the model's context window. Set auto_compact_enabled: true in settings to automatically compress older messages when the token count reaches auto_compact_threshold (default 100,000).
Vault as external memory
For structured, explicit memory — documents, reference files, accumulated knowledge — use the Vault.
Reference vault files directly in an agent's system prompt so they are loaded as context on every request:
# In system prompt
@[memory/user-preferences.md]
@[memory/project-context.md]
Agents can write to vault files using the vault_write or vault_patch tools, building up persistent knowledge bases across sessions.
Available vault tools:
| Tool | Description |
|---|---|
vault_create | Create a new file in the vault |
vault_read | Read a file from the vault |
vault_write | Overwrite a vault file |
vault_patch | Patch a vault file (deep merge for JSON, find/replace for text) |
vault_list | List files in a vault directory |
vault_delete | Delete a vault file |
Native memory tool
Synapse supports Anthropic's built-in memory tool, which lets the model autonomously manage its own memory files — deciding what to store, update, or remove without explicit instructions.
Unlike vault (where you control what gets saved), the native memory tool gives the model agency over its own long-term memory. This is useful for agents that should accumulate knowledge independently across sessions.
The memory tool is available as a beta feature via the Anthropic API. Enable it in an agent's tool configuration to allow the agent to read and write memory autonomously.
Sequential thinking
Synapse includes a built-in sequential thinking MCP server that gives agents the ability to break complex problems into steps before responding. This is available to all agents as an internal tool and does not require explicit configuration.