General agent concepts

Agent memory

Agent memory is what lets an AI agent carry something from one session to the next. A model only sees its current context window, so anything worth keeping, like a user's preference, a project's conventions, or a lesson from a correction, has to be written somewhere else and loaded back in later. As of October 2026 that ranges from memory built into products, such as ChatGPT's memory and the CLAUDE.md files and auto memory notes in Anthropic's Claude Code, to building blocks for developers, such as LangChain's LangGraph long-term store and Anthropic's memory tool.

Memory makes an agent more useful, and it also makes its mistakes last. Claude Code's docs put it plainly: memory is context, not enforced configuration, so a model can still ignore a remembered rule, and a stale or wrong note keeps shaping behavior after the session that wrote it has ended. Microsoft's June 2026 guidance on agent memory safety calls persistent memory a configuration layer for the system, because a memory saved today can change behavior later.

The worst case is memory poisoning: an attacker's instruction or a false fact saved into memory on purpose. The OWASP GenAI Security Project lists it as ASI06, Memory & Context Poisoning, in its Top 10 for Agentic Applications 2026, and 2026 papers showed text hidden in an email or other external content getting agents such as OpenClaw and Claude Code to save an attacker's instruction and act on it in a later session.

8 questions

Answered, plainly.

Can one email change an agent's future behavior?One email can change an agent's future behavior when the agent reads mail and writes its own long-term memory: hidden text gets saved as a fact or rule it reloads and follows later.answer →How do you find and fix a poisoned agent memory?Trace the bad action back to the memory entries the agent loaded and check where each came from, then remove the entry and what was derived from it, and rerun.answer →What is agent memory?Agent memory is information an AI agent saves outside its context window and loads into later sessions, so it can reuse what it learned without being told again.answer →What is agent memory poisoning?Agent memory poisoning is an attack that gets false facts or attacker instructions saved into an AI agent's persistent memory, so they keep steering it in later sessions.answer →What's the difference between agent memory and RAG?Agent memory is a record the agent writes about itself and reads back later; RAG is a fresh lookup against a separate corpus the agent never writes to.answer →How do you test an agent that has memory?Mem0's own tests show a memory lookup can pass while the agent still acts on a stale fact. Check for staleness, contradictions, drift, and cross-user leaks instead.answer →Should an AI agent be allowed to write its own memory?Yes, but not with a blanket write permission: what an agent may save without asking depends on the fact and where it came from.answer →What's the difference between short-term and long-term memory in an AI agent?Short-term memory is checkpointed inside one thread and visible only there; long-term memory is written to a separate, namespaced store any thread can load.answer →

Try Tessary on Cloud.

start for free or self-host

free · 10,000 traces a month · 1 GB · 30-day retention · $10 one-time credit for triage and RCA · no credit card