I have been building AccInt, a local work loop for AI coding agents.
The short version: agents do not just need generic memory. They need a Work Model: a record of the context retrieved, decisions made, failed attempts, tests run, and outcomes that proved whether the work actually landed.
That matters because repeated agent work usually fails in the same places:
- the right context was not retrieved next time
- a past failed attempt was repeated
- passing tests were not connected back to the decision that caused them
- memory grew, but no one knew which memory earned its keep
AccInt is my attempt at making that feedback loop explicit. It uses late-interaction / MaxSim retrieval over scored tokens, commitments and outcomes, and surprise-gated credit so useful context gets stronger only when reality validates it.
I am especially looking for feedback from people using Claude Code, OpenCode, Codex, or building agentic devtools / RAG systems:
- Where do your agents repeat the same mistakes?
- What evidence should count as useful memory?
- What would make a Work Model useful in your workflow?
Early access: https://accint.xyz/
Top comments (0)