I built Lerim, an Apache-2.0 context compiler for AI agents.
Hey everyone. I've been working on an open source project called Lerim, which compiles AI agent traces and truns them into reusable context. v0.3.0 is out now.
The basic idea is simple: completed agent runs should not disappear into raw logs.
When an AI coding agent, support agent, research agent, or incident agent finishes a run, it often leaves behind useful context:
- decisions
- constraints
- facts
- user preferences
- handoffs
- evidence
But the next agent usually starts from scratch or gets a giant transcript pasted into the prompt.
Lerim compiles completed sessions into reusable context records, then makes them queryable through CLI, MCP, and native adapters.
What I added in the latest release:
- Apache-2.0 license
- MCP context tools
- custom YAML source profiles for new domains
- support for custom JSONL traces
- benchmark docs with raw artifacts
- public market comparison docs with source links
Install:
pip install lerim
lerim init
lerim project add .
lerim up
Then:
lerim answer "What should a future agent know before working here?"
Repo: https://github.com/lerim-dev/lerim
Website: https://lerim.dev
I am trying to make agent memory feel less like “store all chats forever” and more like “compile the useful context after each run”.