> For the complete documentation index, see [llms.txt](https://vayl.gitbook.io/vayl-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://vayl.gitbook.io/vayl-docs/documentation/why-vayl/memory-for-ai-agents-that-cant-be-wrong.md).

# Memory for AI agents that can't be wrong

**Vayl is memory for AI agents where a wrong answer has consequences.** Most memory layers *accumulate* — they save every fact and later hand your agent a stale one. Vayl **reconciles**: a new value supersedes the old, "we dropped X" actually removes X, ambiguous input is flagged instead of guessed, and every change lands on a signed, tamper-evident audit trail. It speaks the Model Context Protocol, so any agent — Claude, Cursor, your own — plugs in.

## Why "remembering more" is the wrong goal

An agent doesn't fail because it forgot. It fails because it confidently returned something that *used to be true*. Additive and vector memories keep every version of a fact and let similarity search pick one — so after a few updates, a stale value can rank as "current." The failure is silent: a confident, wrong answer with no signal that it's wrong.

## What Vayl guarantees instead

* **One live value per fact.** A structural invariant — at most one active value per `(subject, scope)` — so "what's true now" is unambiguous. Even a small model can't leave two contradictory values live; the engine won't store them.
* **Removal is first-class.** "We dropped Sentry" *retracts* it. It never comes back as current.
* **The past is still there.** Superseded facts move to history, queryable on demand — not returned by accident.
* **Every change is provable.** A hash-chained, Ed25519-signed audit trail; `verify_audit` pinpoints any edit, reorder, or deletion.

| Property                          | Vayl                                                                        |
| --------------------------------- | --------------------------------------------------------------------------- |
| Silently-wrong on the messy suite | **0%** *(single-run, author-written — reproduce it yourself)*               |
| Cost per remembered fact          | **\~2 LLM calls, no graph database**                                        |
| Storage                           | one **SQLite** file by default; **Postgres** for multi-writer scale         |
| Runs                              | **locally** — no telemetry; the only outbound call is the LLM you configure |

## Not for you if

You need broad document Q\&A over a static corpus (that's RAG), or a knowledge graph for deep multi-hop relationship queries (a dedicated graph reads those faster). Vayl is *reconciling state memory* — it keeps changing facts current, and composes with those tools rather than replacing them.

## Next steps

<table data-view="cards"><thead><tr><th></th><th></th><th data-hidden data-card-target data-type="content-ref"></th></tr></thead><tbody><tr><td><h4><i class="fa-bolt" style="color:$primary;">:bolt:</i> Quickstart</h4></td><td>Store your first reconciled memory in about 5 minutes.</td><td><a href="/pages/416edc1381c20849bc41111b02e482a722a49b94">/pages/416edc1381c20849bc41111b02e482a722a49b94</a></td></tr><tr><td><h4><i class="fa-clock-rotate-left" style="color:$primary;">:clock-rotate-left:</i> Why memory goes stale</h4></td><td>The root cause behind stale answers, and the fix.</td><td><a href="/pages/JBgX4JjZzX9NfXaWCROd">/pages/JBgX4JjZzX9NfXaWCROd</a></td></tr><tr><td><h4><i class="fa-book" style="color:$primary;">:book:</i> Core concepts</h4></td><td>The same-slot invariant and how reconciliation works.</td><td><a href="/pages/FaSGGOl6CSP0t71WWh7l">/pages/FaSGGOl6CSP0t71WWh7l</a></td></tr></tbody></table>


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://vayl.gitbook.io/vayl-docs/documentation/why-vayl/memory-for-ai-agents-that-cant-be-wrong.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.