---
title: "Total Recall | Proactive memory & automated context for AI agents"
description: "Proactive memory and context for AI agents that works on its own: it captures every decision automatically, surfaces the right context the moment it matters, and knows what to keep and what to intelligently forget until it's needed again. 98% on LongMemEval, up to 97% fewer tokens read, zero tokens at rest, 100% local."
canonical: "https://total-recall.dev/"
last-updated: "2026-09-12"
---

# Total Recall

> Proactive memory and context for AI agents that works on its own. Total Recall automatically captures every decision, every why, every how, from both your side and your agent's, then surfaces exactly the right context at the moment it matters. It knows what is current versus superseded, ranks memory by importance, and forgets intelligently, holding superseded facts back until they matter again. 98% end-to-end on LongMemEval, up to ~97% fewer tokens than reloading a full store, zero tokens at rest, 100% local.

Your agent forgets everything when a session ends: every decision re-explained, every fix rediscovered, every project re-pasted into the prompt. Total Recall ends that, with no work from you. Most memory tools make you save, tag, and tweak to keep them useful, and burn tokens doing it. Total Recall flips that: you keep working the way you already do, and your agents just know.

It is not a passive log. It captures both sides of the work as it happens, your intent and decisions and your agent's reasoning, then it manages that memory intelligently: it tracks when a fact changes, demotes the version that has been superseded and surfaces the current one, and ranks what it returns by importance and relevance rather than recency. One shared memory spans your agents, so Claude can recover what you decided with Codex, and the reverse.

## Why it is different

- **Proactive and automatic.** Recall surfaces on its own, before you think to ask. No `add()` calls, no filing, no maintenance.
- **Intelligent, not a flat store.** Total Recall tracks changing facts, demotes superseded ones, and ranks by importance and relevance, so memory stays useful as it grows instead of drowning you in stale hits.
- **Token-efficient at scale.** Zero tokens at rest; a typical recall reads about 700 to 1,700 tokens on a live 10,000+ session memory; up to ~97% fewer tokens read and ~6x lower cost per correct answer than reloading a full store. Cost stays flat as memory grows.
- **Fully local and private.** All conversations, memory, and searches stay on your machine; nothing is uploaded. Search runs locally with zero LLM calls.
- **Cross-agent.** Claude Code, Codex, and Hermes share one memory today, with more agents on the roadmap.

## Benchmarks

Total Recall reports all three questions a memory benchmark asks: did the right memory show up, did it find everything, did the agent answer correctly.

- 98.0% end-to-end judged answer accuracy on LongMemEval (500 questions, gpt-5.4, earlier retrieval snapshot; rerun pending).
- 97.73% Recall@10 (deterministic local retrieval).
- 92.36% Recall@10 on LoCoMo (retrieval-only, v1.9.81).
- On MEME (facts that change), the current fact ranks above the outdated one 99 times out of 100 (retrieval-only).

See the [benchmarks page](https://total-recall.dev/benchmarks) for full methodology.

## Pages

- [Home](https://total-recall.dev/)
- [Use cases](https://total-recall.dev/use-cases)
- [Benchmarks](https://total-recall.dev/benchmarks)
- [About](https://total-recall.dev/about)
- [Contact](https://total-recall.dev/contact)
- [Privacy](https://total-recall.dev/privacy)
- [License](https://total-recall.dev/license)

## Get started

Total Recall is launching in public beta. The public beta is free; after that, pricing will offer a generous free tier plus paid and enterprise tiers.

- Request early access: https://total-recall.dev/ (join the waitlist)
- Questions or beta access: hello@total-recall.dev