// the find
CaviraOSS/LongMemory
Local persistent memory store for LLM applications including claude desktop, github copilot, codex, antigravity, etc.
LongMemory is a TypeScript memory engine for LLM apps and coding agents. It stores memories in a local SQLite file and recalls them through strict, historical, associative, and world-grounded modes, exposed as an npm library, a CLI, an HTTP and MCP server, and plugins for hosts like Claude Code and Gemini CLI. It suits developers who want agent memory they host and can inspect, rather than a hosted memory API.
Recorded time and valid time are stored separately, so a question like 'what was the deployment region in January' is answered from history with superseded facts kept, not from a filter on the latest row. The SQLite default is a sensible choice: one file, reopening restores nodes, edges, temporal history, and lifecycle state, and nothing else needs to run. Server-bound identity is the right call for an MCP surface. Tool arguments cannot override user, project, or agent scope, and the README tells you to treat recalled content as untrusted evidence rather than authorization.
The surface is large for what the core has to prove. The README lists a library, CLI, HTTP server, MCP server, Next.js dashboard, VS Code extension, n8n node, Python client, session porter for several coding agents, cloud document connectors, and integration packs for about fifteen hosts. Each one is a maintenance commitment, and the README does not say which are stable. The benchmarks are the part that would settle the quality question, and the README cites LongMemEval, LoCoMo, and BEAM without showing a single number. You have to run pnpm bench yourself, and official scorecards refuse to run on embedding fallback, so you need a real embedding provider to get a meaningful result. 'Hydrograph' is the name of the core graph substrate and is never defined in the README, so the central abstraction is left for the reader to find in the architecture docs.