Compact MCP memory engine for long-term agent context
dm-lite (daimon-memory v2) by Wakbijok supplies persistent memory for AI agents, operating as a Model Context Protocol server to store and recall long-term context across sessions. The tool combines keyword matching with vector embeddings to improve retrieval, offers bitemporal history for past-state queries, and delivers a small, typed binary with multitenant support. Developers building autonomous agents for localization, coding, or text processing gain a compact memory engine with offline deployment options.
Designed to plug into MCP agent workflows and store session-to-session state
The tool runs as an MCP-compatible server and integrates with agent hosts that speak the protocol. It is distributed as a cross-platform Rust binary and requires an MCP-compatible host plus a Node.js environment for certain integration hooks. Supported agent examples include Claude Code, Devin, and Claude Desktop. The server accepts networked endpoints or a local loopback endpoint, making deployment flexible for desktop and networked agents.
Retrieval quality depends on embedding model choice and index placement
dm-lite stores dense vector embeddings alongside typed SQLite storage, so retrieval behavior reflects the embedding model used and the index layout. The default embedding model footprint is about 130 MB on disk, and each tenant maintains its own vector index and SQLite file. Practical result: query relevance improves with higher-quality embeddings and careful tenant indexing, while noisy or undersized models reduce hit rates.
Fits compact, developer-focused deployments rather than large vector clusters
The project emphasizes a low operational footprint compared with cluster-style vector databases, making it suitable for research, prototypes, or single-host agents. Multitenancy creates isolated data pools for separate agents, and the embedding cache can be pre-populated for air-gapped installs. The project is actively maintained within the MCP developer community, which helps with compatibility updates and protocol changes.
Practical choice for MCP developers who need compact, maintainable memory
dm-lite is a pragmatic option for developers building MCP agents who want local, maintainable long-term context and an actively maintained project. Teams prioritizing very large-scale, enterprise vector indexing should evaluate dedicated cluster databases instead. The tool suits prototypes, desktop agents, and research use where a small, hostable memory engine is preferable to a full vector cluster.
