rememori

rememori is an embedded memory engine designed for AI agents, implemented in pure TypeScript with zero dependencies, zero servers, and zero native bindings. It addresses the challenge of providing persistent, in-process memory for agents that would otherwise lose context, user preferences, and prior interactions at the end of each session. The tool offers a lightweight alternative to external vector databases and retrieval services, enabling agents to remember and recall information efficiently without additional infrastructure.

The core API revolves around four main operations: remembering, recalling, forgetting, and reinforcing memories. Text can be stored along with its embedding, tags, and automatically extracted entities. Recall operations rank stored memories using a combination of similarity, shared entities, user feedback, importance, and time decay, while the forget function removes specific memories. The reinforce function strengthens memories only when there is verifiable evidence of their use, and demote provides a mechanism to reduce the prominence of incorrect memories. rememori automatically constructs a bipartite knowledge graph by linking memories to their mentioned entities, enhancing recall accuracy even when queries do not directly match stored text.

Developers can use any embedding model, including local options like Ollama, OpenAI-compatible endpoints, or custom functions. Storage is handled locally: on Node and Bun, data is kept in a single append-only file; in browsers, IndexedDB is used; and on edge platforms, a key-value adapter is available. The tool operates wherever JavaScript runs, including Electron and edge workers, and it is suitable for a range of applications such as chatbots, coding agents, local-first apps with semantic search, compliance-sensitive assistants, home automation, and game NPCs.

rememori is distributed as a single npm package, with a core size of approximately 8 kB gzipped, and is licensed under the MIT license. It is not intended as a multi-user server, cloud solution, or document-ingestion platform, but rather as a memory engine embedded directly within an agent's process for rapid deployment and minimal operational overhead.