Pensyve
A framework-agnostic memory runtime that lets an AI agent store and later recall facts, episodes and what worked before, using local embeddings by default or a Postgres backend for production scale.
Pensyve is a universal memory runtime for AI agents, built to work offline first and independently of any particular agent framework. It stores three kinds of memory: episodic memory of what happened, semantic memory of what is known, and procedural memory of what actually works, the last of which tracks action-to-outcome pairs so an agent stops repeating an approach that failed before.
Retrieval combines eight signals, including vector similarity, keyword search, graph relationships, intent, recency, frequency, confidence and memory type, and a forgetting-curve mechanism lets memories that keep getting used grow stronger while unused ones fade, with repeated facts consolidated over time. By default it runs offline with a local SQLite database and local embeddings, needing no API key; a Postgres backend with a vector extension is available for multi-node production deployments, with a managed hosted option also offered.
It provides libraries for Python, TypeScript and Go, an MCP server, a REST API and a command line, plus adapters for LangChain, CrewAI and AutoGen.