Memory that grows with the user. And stays correctable.
For journaling apps, coaching bots, and AI companions, memory is the product. MemexAI stores what the agent learns as inspectable files — so users can see exactly what the AI knows about them, and correct it when it's wrong.
A hidden memory profile is a liability.
The promise of a personal AI assistant is that it knows you — your goals, your preferences, your context. That promise only holds if the memory is accurate. When it isn't, users notice immediately. And unlike a search ranking or a recommendation feed, a personal AI speaking with confident incorrectness feels like a betrayal.
With embedding-based memory, the user has no recourse. They can't see what the assistant learned. They can't tell if it noted something correctly or misinterpreted a message. They can't fix a stale fact without explicitly re-explaining themselves — and even then, the old entry stays in the index, silently competing.
Trust erodes quietly. The user corrects the AI once. It happens again. They stop engaging with topics the AI seems to have wrong. Retention drops not because the product is slow or buggy — but because the memory can't be trusted.
The deeper problem is that memory without a correction surface is opaque by design. There's no admin panel a support person can open. No revision to roll back to. No way to tell whether a bad response came from wrong memory or a bad retrieval.
For a product where memory is the moat, that opacity is the exact opposite of what you need.
Memory the user can see. Memory your team can fix.
MemexAI stores everything the agent learns about a user as path-addressable Markdown files in your own Postgres. Not embeddings. Not a managed cloud API. Files you can open, read, and edit.
Integration
Call getPromptBlock to inject the user's current memory into the system prompt before every model call. Wire memory_remember and memory_context as tools so the agent can write durable facts and retrieve relevant context as it works.
import { createMemex } from '@memexai/core'
import { generateText } from 'ai'
const memex = createMemex(DATABASE_URL)
// Load everything the agent knows about this user
const { block } = await memex.getPromptBlock({
userId: session.userId,
})
const { text } = await generateText({
model,
system: `You are a personal AI assistant.\n\n${block}`,
messages,
// memory_remember: write a durable fact
// memory_context: retrieve relevant context
tools: memex.tools({ userId: session.userId }),
})What the user gets
Memory files are human-readable Markdown. A user asking "what do you remember about me?" gets an honest answer — because the answer is literally a file they can be shown. Your support team can open any user's memory in the admin console, see every revision with its timestamp and tool call ID, and edit wrong facts directly.
The shared/ namespace holds global guidance — tone instructions, product policies, schema definitions for what a well-structured user profile should contain. When you update the profile schema, all future agent sessions reflect it without a deployment.
Background dreaming compacts and cleans memory files between sessions — merging duplicate notes, removing stale entries, keeping the profile accurate as the user evolves.