Top AI Product

Every time a Claude Code session ends, the agent forgets your architecture, naming conventions, and past decisions — so you re-explain the project from scratch next time, burning tokens and patience. Recall, an open-source tool that hit Show HN, fixes that by giving Claude Code durable memory that lives entirely on your machine.

## What Recall does

Recall plugs into Claude Code via the Model Context Protocol and persists project context across sessions, so the agent can look up what it already learned instead of re-onboarding every time. The author reports 94.5% LoCoMo recall@10 — a benchmark for long-conversation memory retrieval — at roughly 70ms p50 latency, which is fast enough to query inline without slowing the agent down.

## Local-first, no API keys

The selling point is that it’s fully offline: no external service, no API keys, your code and context never leave the machine. That’s a meaningful contrast with hosted memory layers, especially for developers who can’t ship proprietary code to a third-party vector store. It joins a fast-growing class of “agent memory” tools, but the local-first, zero-dependency framing is what’s drawing attention.