Recall

Local-first agent memory — from raw logs to distilled wisdom, with semantic search over all of it. Comes with Recall Bench, a benchmark harness for evaluating any agent memory system (Recall, OpenClaw, Loki, your own) against a synthetic multi-year persona corpus.

This site is organized into two sections so you can dive into the one you care about and ignore the other:

Recall Memory System →

Architecture, the four-level compaction pipeline (daily → weekly → monthly → wisdom), the two-phase hierarchical search, dreaming and contradiction detection, and the prompts that drive it. (Recall is an in-progress, experimental system — see the note inside.)

Recall Bench →

The benchmark itself — the scoring dimensions and ten recall categories, the persona corpus, harness adapters, how to run it with a coding agent, and published runs and heatmaps for each memory system tested.

Recall is an open-source project. Source: github.com/Stevenic/recall.