Synapse

What Synapse does

Synapse extracts semantic facts from episodic conversations and stores them in local SQLite. Facts with a single current value are anchored to entity keys, so a new value supersedes the old one. The system detects and resolves conflicts, decays stale information, and refuses to store credentials. An always-on digest holds the facts an agent should know at the start of a session.

Retrieval is a hybrid score that combines vector similarity, BM25 keyword matching through SQLite FTS5, importance, confidence and recency, minus decay and conflict penalties. Retrieval does not call an LLM by default; the LLM is used only for consolidation and conflict detection, and an optional rerank flag lets an LLM reorder the final results. Agents connect through a stdio MCP server or SDK adapters. Users get a CLI with remember, query, export and decay commands, an inspector page, and export and purge controls. The monorepo is written in TypeScript.

Key features

• Typed long-term memory on local SQLite, with entity-key supersession of outdated facts
• Hybrid retrieval scoring vector, BM25 keyword, importance, confidence and recency
• Conflict detection, decay of stale facts, and user-controlled export and purge
• Stdio MCP server, HTTP API built on Hono, CLI and an inspector UI
• Feedback tool that marks memories as helpful, stale or wrong

When to use it

• Giving an agent memory that persists across sessions without a hosted database
• Keeping a changing fact, such as a current preference, from clashing with its older value
• Exporting or purging everything an agent has remembered about you

Who it is for: Developers who want a local, inspectable long-term memory layer for AI agents and are comfortable wiring up an MCP server or API.

How it fits with Hermes Agent

The repository is tagged hermes-agent and nous-research on GitHub. The README documents a stdio MCP server and lists Claude Code, Claude Desktop, Cursor and OpenCode as example clients, so Hermes is one of several agents it can serve.

FAQ