v0.1.0 is out
Codebase memory for coding agents.
Your agent gets only the code it needs, within a token budget. Open source and local first.
npx -y remimem init claude mcp add remimem -- npx -y remimem serve
86% (86%)
fewer tokens per task than an agent that greps and reads whole files, on 36 held-out tasks from real commits.
0.93 (0.93)
recall@10 on held-out tasks
35 ms (35 ms)
p95 query latency
38 s (38 s)
to index 99k lines
How it works
Index once. Ask anything. Remember what matters.
01
Index
Parses your code into functions, classes, and the calls between them. Locally, in seconds.
02
Retrieve
Finds the code that answers the question by name, by meaning, and through the call graph.
03
Remember
Keeps the decisions your team made, linked to code, and flags them when the code changes.
Benchmark
Read the benchmark Measured on hono, httpx, and echo. Method, every table, and what the numbers do not show.
Local
Parsing, embedding, and storage run on your machine, in one SQLite file. No telemetry.
Budgeted
Every answer fits the token budget you set and says exactly what it left out.
Open
MIT licensed. Every number on this page comes from a benchmark you can rerun.
Give your agent a memory.
npx -y remimem init