recall reads the chats Claude Code, Codex and Cursor already keep on your machine. Every agent gets search across all of them, and notes that update overnight.
Coding agents forget everything between sessions, and none of them knows what you told the others.
recall gives every agent two things:
It runs locally. Search needs no model at all, and you choose the model that writes the notes.
recall doesn’t capture anything new. It reads the histories already on disk, read-only: each chat is split into your turns and the agent’s, tool output and injected text are dropped, and secrets are redacted before anything is stored.
Clones and worktrees of one repo count as one project, keyed on its git remote.
Search is a full-text index on your machine, like the index at the back of a book: every chat with a message that matches a query word is a candidate (Figure 2). No model is involved, so a query takes 6–20 ms.
recall then ranks those matches five ways and fuses the rankings:
Once a night, recall turns new chats into notes (Figure 3). Each run is one git commit, so you can see what changed and undo it.
Notes are plain markdown. Edit or delete one and the next run keeps your change.
A project is rarely one folder. It’s a repo cloned twice, a few worktrees, the path it had before you moved it, and chats with no folder at all.
recall follows all of them as one project (Figure 4), so an agent that starts in any of them gets the same notes. It also flags rules one agent reads and another never sees, like a Cursor rule Claude Code has never loaded.
Measured on LongMemEval and LoCoMo, with each benchmark’s own prompts. Every number reproduces with recall eval.
Given only the sessions that hold the answer, the same Haiku model scores 92.0%, so imperfect retrieval costs about five points.
Reading more closes most of that gap (Figure 5). With the top 20 chats, Haiku reaches 90.5% for under a cent a question, more than Sonnet 5.5 reading excerpts at a quarter of the cost.
Published LongMemEval_S scores run from 71% to 95%. Each system reports its own run, with its own answer model and judge, so read the ranking loosely.
One difference does compare cleanly: other memory systems call a model for every message they store. recall calls none.
recall hands the answering model about 44k tokens a question. Mem0 hands it about 7k. The gap shows most on preference questions and on questions that span several sessions.
Closing it without a model call per stored message is the next piece of work.
Let your agent do it. Paste this into Claude Code, Codex or Cursor:
Install recall for me: follow https://raw.githubusercontent.com/alirng/recall/main/INSTALL.md
Or by hand. macOS or Linux, Python 3.11 or later:
uv tool install 'recall-agents[embeddings]' recall setup
Setup finds your agents, asks which model should write your notes, builds the index and schedules the nightly run.