What do people use to give an agent memory across sessions?
Every new session my agent starts from zero. What are people actually using so it remembers facts, preferences and past decisions between sessions, and does a memory server really help?
Most working setups use two layers. Project facts live in plain files in the repo (notes, plans, a rules file), because any tool can read them and you can review them in git. Personal and cross-project facts go into a memory server the model can read and write.
The reference Knowledge Graph Memory server is the simplest place to start: it stores entities, relations and short observations in a local file, and any MCP client can query it at the start of a session.
The storage is the easy part; deciding what to save is hard. Without rules, memory fills up with stale or wrong notes that then mislead the agent. Save decisions and corrections, not chat logs, and prune what no longer holds.
Measure it on a real task, as the original post did: run the same job with and without memory and compare. If the agent never reads what it saved, the memory is only costing you context.
Listings mentioned
- Knowledge Graph Memory Server · agent by modelcontextprotocolReference MCP server: a local knowledge graph the model can read and write.
- Memory curator · agent by VoltAgentRules for what to save, what to recall and what to prune over time.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.