the agent memory problem have just been solved. it’s free and easy to use.

the project is called mempalace.
the idea is simple:
most memory systems try to decide what matters, summarize it, and store only that.
what mempalace does:
store everything, then make it searchable.
instead of extracting stuff like:
• “user prefers postgres”
• “project uses graphql”
and throwing away the conversation where the actual reasoning happened,
it keeps the original exchanges verbatim and uses semantic search to retrieve the right memory later.
so you keep the why, not just the summary.
how it works:
> everything stays local
> it stores your data in ChromaDB
> it organizes memory into a structure they call a “palace”
> then your agent can search that memory when needed
how to use it:
1. install mempalace
2. initialize it on your project
3. mine your chats / docs / codebase
4. connect it to your AI through MCP
5. from that point, your agent can query your past decisions automatically
they also claim 96.6% on LongMemEval in raw mode, locally, with zero API calls.
so is basically:
• local memory
• no cloud
• no subscription
• no summarization destroying context
• just full recall when your agent needs it
honestly, this is the right direction for agent memory.