RE-callSetup
English source

Setup guide

See it work in two commands.


A local memory layer for agents. The second command starts its own database, indexes a sample corpus and answers three questions, one of which it refuses.

Terminal
pip install "recall-rag[fastembed]"
recall quickstart

Needs Python 3.11+ and Docker. Nothing is calibrated and nothing is wired to an agent: it prints the next command for each. recall quickstart --remove deletes everything it created.

Python 3.11+ PostgreSQL 18 via Docker Windows · macOS · Linux Runs offline
ATM-Bench Recall@10 92.89 vs 79.09 best published (limits) MTRAG correct refusals second of ten (limits) Zero memory-layer LLM calls to build memory (limits)

What you are installing

A memory that would rather say nothing than guess.

RE-call stores notes, decisions, and project history, then searches them by meaning.

When nothing matches with enough confidence, it abstains instead of handing you the nearest note.

The middle question recall quickstart asks

Output
query='how many requests per second can a client make?'
  ok           conf=1.00  cos=0.784  rate_limits_v2.md
  superseded   conf=1.00  cos=0.806  rate_limits_v1.md -> use rate_limits_v2.md

Copied from a real run against the sample corpus. Look at the cosines: the stale memory scores higher (0.806 against 0.784), so plain vector search returns the limit that no longer exists. RE-call flags it superseded, points at its successor, and puts the current memory on top; the third question, which the corpus cannot answer, is refused. Until you fit a threshold to your own corpus every line also carries DEGRADED, which is what calibration fixes.

You need
Python 3.11+3.12 to 3.14 also supported
And
Docker DesktopRuns the database for you
Your data
Stays localThe default embedder runs on your machine
Licence
Apache 2.0Free for commercial use

The moving parts

Three things, and only one of them is RE-call.

Most of the confusion in a first install comes from not knowing which piece a command is talking to. There are only three.

PieceWhat it doesYou touch it
A PostgreSQL database Stores your notes and their vector embeddings. Once, during setup.
The RE-call package Indexes markdown and decides whether a result is safe to use. Every query.
The MCP server Lets Claude Code query your local memory. Optional.

New, and in beta

On Windows, a wizard can do the whole install for you.

One window asks a short list of questions, then brings up the database, indexes your folders and writes the Claude Code wiring while you watch the log.

It is a beta and it is not code signed yet, so the download page says plainly which machines Windows will let it run on. I am looking for testers, and for comments on how it went.

What you give up by using it

Nothing that cannot be recovered. The wizard writes the same configuration file the terminal flow writes, and the install it produces is one you can re-run, inspect, or hand to somebody else.

The full install stays the path on macOS and Linux, and the path to fall back to if the installer will not start.

Start at step one.

Install two things, then verify each step as you go.

Start the install