Agents that learn from every run.
Memspire is the open-source learning layer for AI agents. It turns every graded run into a lesson, distills the best lessons into skills, and ships fixes to your repo as pull requests, so your agent is better on day 10 than on day 1.
Open source · Apache-2.0 · launching soonThe arm that learns. Every run is wound in; what helps stays close, and what misleads is let go.
From a mistake to a lesson it keeps.
On day 1 a refund agent refunds the wrong charge, and you say why. By day 10 it starts the same task briefed with what it learned, and gets it right. Advice that helps rises; advice that misleads is archived, never deleted.
Lessons
- Learned
Check the ledger, and refund only the duplicate charge.
From your review of run #1 · L-104 - Proven
Retry refunds with the same idempotency key.
Followed in 3 runs · score 0.74 · L-117 - Archived
Contradicted twice · archived with a reason · L-098Always refund to store credit. - Learned
Send refunds over $500 to finance for approval.
From a failed run · run #188 · L-121
Illustrative interface with fictional data
What changes when agents learn.
Say it once. The next run starts with it.
Each correction becomes a lesson your agent is briefed with, not a fix that lasts one run.
Lessons arrive only when they apply.
A few at a time, under a token budget. When nothing fits, nothing is added.
You see what a change fixed, and what it broke.
New versions are replayed on your recorded runs, then ship to your repo as a pull request you review.
Retrieval is not memory.
Agents that only search documents don't get better with use. What the last year of research shows about how memories should be written, updated and forgotten, and what to do in production.
Stop re-teaching your agents.
Memspire is open source under Apache-2.0 and runs on your own stack, with your keys and your data. Get it the day it's public, or tell us about the agent you want to improve.