What is agentic memory?
What is agentic memory? It is memory an agent writes and reads itself through tools, so a fact can survive past one chat. In XERJ the agent calls xerj_memory_store when a fact is worth keeping, and xerj_memory_recall when it needs it back. Nothing is saved unless the agent asks. The how-to lives on the agentic memory answer.
TL;DR Agentic memory is memory an agent writes and reads itself through tools. A fact can survive past one chat, because the agent chooses when to store it and when to recall it.
Three memory designs
The tool writes on its own
Some tools save session text or model-written summaries through hooks, without the agent choosing each fact. claude-mem and MemPalace work this way, in different forms.
A store call, then a recall call
In XERJ the agent calls xerj_memory_store when a fact is worth keeping, and xerj_memory_recall when it needs it back. Nothing is saved unless the agent asks.
Memory plus the rest of your data
The same local node can also search logs, docs, code and past session files. Namespaces isolate by name, not by credential.
Act as a coding agent. Read https://xerj.org/llms.txt, install XERJ, start a local node, index one folder I point you to, and run one search so I can see a cited passage.
How it works
A chat alone forgets when the session ends. Some tools fix that with automatic hooks: they write session text, or a model-written summary, without the agent deciding each fact.
Agentic memory works the other way. The agent makes a store call when a fact is worth keeping, and a recall call when it needs that fact back. The agent is responsible for what is saved.
How XERJ uses it
In XERJ the tools are xerj_memory_store and xerj_memory_recall. Nothing is saved unless the agent asks. The same local node can also search logs, docs, code and past session files. Namespaces isolate by name, not by credential.
The how-to is how to use agentic memory in XERJ. A published comparison, read against those projects' READMEs on 2026-10-30, is claude-mem vs MemPalace vs XERJ. This page does not repeat that comparison's percentages; they stay there.
The tools are reached over the Model Context Protocol. A related MCP memory how-to is what is an MCP memory server.
History and research
The product split above, and the comparison page, are the sources for this entry. This page does not attach an extra paper to the name "agentic memory."
Guides
Set it up with the agentic memory recipe. Then read what is an MCP memory server and the AI agent memory benchmark.
Related terms
Model Context Protocol and context engineering.
FAQ
What is agentic memory?
It is memory the agent itself chooses to store and recall through tools, so a fact can outlive a single chat.
How is this different from automatic hooks?
Some tools save session text or summaries through hooks, without the agent choosing each fact. Agentic memory waits for an explicit store call.
How does XERJ store a memory?
The agent calls xerj_memory_store, and later xerj_memory_recall. The same local node can also search logs, docs, code and past session files.
Where is the how-to?
The practical guide is the agentic memory answer. A side-by-side with claude-mem and MemPalace is the comparison page. This science page is the concept.
Related
- How to use agentic memory in XERJ
- MemPalace vs claude-mem (and XERJ): which memory should Claude Code use?
- What is the Model Context Protocol?
- What is context engineering?
- What is an MCP memory server?
- BM25 vs vector search · same intent: Science
- What is Corpus Hub? · same intent: Science
- What is extractive answering? · same intent: Science