Synapse Layer is open-source persistent memory infrastructure for AI agents and assistants, positioned as the missing layer between stateless large language models and real contextual continuity. It encrypts memories at rest with AES-256-GCM, uses pgvector HNSW to build semantic indexes to support semantic recall, and exposes interfaces via MCP JSON-RPC, allowing native integration with Claude, GPT, Gemini, and any MCP-compatible client. The project uses the Apache 2.0 license.
Primary uses It solves the problem of AI agents forgetting everything between sessions, providing agents with long-term memory capabilities across sessions.
Main capabilities
- Long-term assistant memory: persistently preserve user preferences, facts, and prior decisions across sessions.
- Cross-agent continuity: save context in one agent and recall it in another.
- Secure memory for MCP clients: connect MCP-compatible tools such as Claude Desktop and Cursor to a governed memory layer.
- Team operational memory: maintain structured context, trust scores, and retrievable recall for production agents.
- Currently exposes 13 MCP tools, covering memory capture, semantic recall, structured storage, feedback loops, agent handover, and runtime observability, such as save_to_synapse, initialize_context, memory_feedback, neural_handover, etc.
How to use
- Installation and onboarding: obtain a Connect Token at forge.synapselayer.org, then install it into any MCP client; the Smithery card in the Forge dashboard copies the command with the token already embedded (… | bash -s -- sk_connect_…), and pasting it once in the terminal is enough.
- Python SDK: access Forge memory directly from applications via the Python SDK, suitable for native Python workflows and quick integration into existing applications, prototypes, and scripts.
- Cloud / Forge API: use when persistent, cross-session, cross-agent memory and managed access tokens are needed, suitable for production assistants, multi-agent systems, MCP-based integrations, and shared memory across tools and sessions.
- MCP integration: add the configuration to claude_desktop_config.json. Configuration file locations: on macOS at ~/Library/Application Support/Claude/claude_desktop_config.json; on Windows at %APPDATA%\Claude\claude_desktop_config.json; on Linux at ~/.config/Claude/claude_desktop_config.json.
- Authentication uses Header-First Auth; for vulnerability reports, see SECURITY.md.
Applicable scenarios Suitable for long-term assistants that need memory across sessions, multi-agent systems, production agents, and teams that want to provide governed shared memory for MCP clients such as Claude Desktop and Cursor.
Limitations and notes All public claims follow the Public Claims Matrix; architectural details that reveal benefits are public, while the mechanisms that implement those benefits are not public. The project follows the principle of "claims are reality": features that are not implemented will not be written into the README. Specific pricing, fees, platform coverage, or performance results are not yet provided in the materials.
