Wink Pings

AI Agent Memory Packed Into a Single File—So What Should It Forget Next?

AgentDB packs vectors, indexes, feedback state, and audit logs into a single .rvf file, eliminating the need for a separate database server. It makes persistent memory easier to deploy and makes the stale-memory problem more visible.

Dan Kornas has released an open-source project called AgentDB, with a very specific goal: give AI agents persistent memory and retrieval without running a separate database server.

This is familiar to anyone building agents. To let an agent reuse past tasks, results, and context, you usually need to connect a vector database, maintain indexes, handle retrieval results, and then figure out how to record which results were actually useful. AgentDB brings these capabilities into a local memory store. It is open source under the MIT license.

A few of its design points:

  • Single-file storage: vectors, indexes, learning state, and audit logs live in a single .rvf file
  • Feedback-driven retrieval: record which retrieval results are useful and let the learning loop adjust ranking
  • Hybrid search: BM25 keyword matching plus dense vector retrieval, covering both exact terms and semantics
  • Episodic memory: store tasks, results, and evaluations, then retrieve relevant historical episodes later
  • MCP integration: expose memory tools to MCP-compatible clients such as Claude Code and Cursor, so you don’t have to build your own tool interface

AgentDB tweet image

In terms of positioning, AgentDB suits developers who want to quickly try persistent memory, as well as people who don’t want to maintain a database service for a small project. Single-file storage and MCP integration lower the integration cost—at least you don’t have to stand up a whole infrastructure before validating whether memory is useful.

But with memory, storing it is only the first half. Gao Yang mentioned in a reply that local storage is appealing to him; the reason he stopped using a hosted memory service was that it once surfaced an old preference from months ago, and the agent acted on it. What he cares about is pruning: how to decide what should be forgotten. His own notes file keeps growing, and retrieval gets noisier every month.

Thomas Wight’s feedback was more direct: storage was never the bot’s problem—expiration rules are. He mentioned that he set a model rule for a bot at 3 p.m. on October 31, and by 3:26 p.m. it had to be changed. Now the rules carry dates, and the newest one wins. He asked how AgentDB knows when a memory has expired.

These two questions point to the same place. AgentDB has feedback records and audit logs, so it can track which results were used and which were useful, but the "forgetting policy" still has to be defined by the user. Feedback can adjust ranking, and audits can trace provenance, but expiration judgments won’t appear automatically.

One commenter joked that there’s finally an agent that remembers yesterday’s mistakes. Behind the joke is reality: many agents don’t lack storage; they lack the ability to say no to old information. AgentDB lowers the deployment barrier by a notch, and it also puts the forgetting problem front and center.

GitHub: https://github.com/ruvnet/agentdb

发布时间: 2026-11-17 11:30