An AI agent without memory repeats its mistakes and forgets its decisions. What persistent memory for agents really needs (meaning search, the latest decision winning, time, conversation history, isolation), and how to give any MCP agent a hosted memory with one address.
An agent that forgets is an agent you babysit. It asks the same questions every run, retries the approach that failed yesterday, and drops the decision you made last week.
Context windows got big, but a context window is not memory. It is gone when the session ends, and it cannot span your other tools. Persistent memory is what survives.
What agent memory actually needs #
A table of notes is easy. Notes that stay useful after months are not. In our experience running our own work on agent memory every day, five things matter.
1. Search by meaning. The agent will never ask with the same words it used to save. "Which database did we pick?" must find "Decided: Postgres, because orders and stock need one transaction". Keyword search alone misses it.
2. The latest decision wins. Plans change. If the old decision and the new one both sit in memory with equal weight, sooner or later the agent acts on the old one. Updates must replace, with history kept for "what did we decide before?".
3. Time. "What did we do last week?" and "when did the plan change?" are among the most common questions. Memory needs dates and a timeline, not only similarity.
4. Conversation history. Facts are distilled. Sometimes the agent needs the exact moment: what the user said, which step it stopped at. Keeping the conversation, searchable, makes "continue yesterday's work" possible.
5. Isolation and safety. One user's memory must never leak into another's. Secrets like API keys must never be stored. The user must be able to see, export and delete everything.
Add it to any MCP agent #
NestDaddy Memory is a hosted memory server over MCP that covers all five. Your agent gets tools to:
recall: given the user's message, return the memories and past chat moments that clearly relate (call it on every turn),save_memory,search_memories,get_memory: one topic per memory, found by meaning in any language,memory_timelineandtimeline_events: how a decision changed, and what happened in a period,log_turn,search_conversations,read_conversation: keep and find past conversations,find_duplicatesandmerge_memories: keep it tidy as it grows.
The endpoint is https://nestdaddy.com/memory/mcp (Streamable HTTP, OAuth 2.1 with discovery). Any MCP client that supports remote servers can connect: it opens a sign-in page, the user clicks Allow, and there are no keys to manage.
In Claude Code, for example:
claude mcp add --transport http nestdaddy-memory https://nestdaddy.com/memory/mcp
Then authenticate from /mcp. In Cursor, VS Code or Windsurf, add a remote MCP server with the same address.
Tell the agent how to use it #
The server sends usage instructions to the client, and a short standing rule makes it consistent:
On every message, first call recall with what the user said. Before deciding anything about a project, search memory. When we decide something, save it (update the existing memory instead of adding a duplicate). After each answer, log the turn.
One memory, every agent #
Because the memory sits outside any one app, every agent and AI app the user connects shares it. A decision made in ChatGPT on a phone is there for the coding agent on the laptop. Switch models or tools and nothing is lost.
Start today #
Memory compounds: every run from today makes the next one smarter. Start your free month, no card: nestdaddy.com/memory.