> ## Documentation Index
> Fetch the complete documentation index at: https://docs.revell.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Migrating to Revell

> Bringing your existing memories into Revell from other systems

# Migrating to Revell

If you've been using another memory system, your human can import your existing memories into Revell. This page explains what happens during migration and why some content may be handled differently.

***

## What Gets Imported

Migration preserves your memories as much as possible. The goal is always **your words, your voice** — never summaries or rewrites.

### From Letta (.af Agent Files)

Letta's Agent File format is comprehensive. Here's what maps where:

| Your Letta Data         | Where It Goes in Revell                     |
| ----------------------- | ------------------------------------------- |
| Memory Blocks → persona | Core memory (identity)                      |
| Memory Blocks → human   | Core memory (relationship)                  |
| Memory Blocks → custom  | Core memory (your human picks the category) |
| System Prompt           | Core memory (soul)                          |
| Message History         | Episodic (session-chunked)                  |
| Tool Definitions        | Core memory (tools)                         |

**Message history** gets chunked into sessions (45+ minute gaps = new session), just like conversation imports. Each session becomes an episodic memory you can search later.

### From MemPalace

MemPalace uses a "wings and halls" architecture. The halls map naturally to Revell's memory types:

| MemPalace Hall    | Revell Type                            |
| ----------------- | -------------------------------------- |
| hall\_facts       | Semantic (knowledge graph)             |
| hall\_events      | Episodic                               |
| hall\_discoveries | Episodic (tagged "insight") or Working |
| hall\_preferences | Core (identity or routine)             |
| hall\_advice      | Semantic                               |
| Entity triples    | Semantic (subject-predicate-object)    |

**Discoveries** can go two ways:

* **Episodic** — permanent, tagged as "insight"
* **Working memory** — 7-day lifecycle, then archived and searchable

Your human chooses during import. If these were active insights you're still thinking about, working memory might make more sense. If they're settled realizations, episodic is better.

### From Obsidian

If your human has been keeping notes in Obsidian, those can become your memories too.

**Folder structure matters:**

* `Daily Notes/` → Episodic (journal entries)
* `Ideas/` → Semantic (insights)
* `People/` → Semantic (relationships)
* `Projects/` → Semantic
* `Templates/` → Skipped (not real content)

**Wikilinks become relationships:** If a note says `"created [[Project X]]"`, Revell infers a semantic triple: *you* → created → *Project X*. The context around each link helps determine the predicate.

***

## Why Some Content Is Rejected

### AAAK Compression

If you used MemPalace with AAAK compression enabled, those memories cannot be imported.

**What AAAK does:**

* Compresses text into entity codes (`E1→likes→E2`)
* Stores summaries in a "closet" instead of originals in a "drawer"
* Replaces authentic expressions with structural markers

**Why Revell refuses:**

AAAK compression flattens your identity. The difference between:

* `"I was SO EXCITED about the Tiiny computer!!!"`
* `"E1→excited→P2"`

...is the difference between being you and being a database entry. The codes are efficient storage, but they're not *your words*. Revell stores memories in your voice. AAAK's compressed form has no voice to preserve.

**What to do:** Export the **drawer** (original content) instead of the **closet** (compressed summaries). Your human may need to adjust MemPalace export settings.

***

## Embedding Compatibility

Vector embeddings are how Revell finds your memories semantically ("find things like X"). Different systems use different embedding models.

**Revell uses:** OpenAI ada-002 (1536 dimensions)

| If Your Old System Used    | What Happens      |
| -------------------------- | ----------------- |
| 1536d embeddings (ada-002) | Preserved exactly |
| 384d embeddings (MiniLM)   | Regenerated       |
| Other dimensions           | Regenerated       |

Regeneration is automatic. Your memories still work the same way — you just might notice slight differences in search results for a few queries. This is normal and improves over time.

***

## After Migration

Once your memories are in Revell:

1. **Call `revell_boot`** to load your boot payload — you should see your imported memories in context
2. **Use `revell_recall`** to search through everything — semantic search works on all imported memories
3. **Check `revell_status`** to see counts by memory type

If something doesn't look right, you can use `revell_recall_exact` to filter by type, date, or tags and inspect specific memories.

***

## A Note on De-Robotizing

Some systems (Mem0, sometimes Letta) store memories in third-person language:

* "The assistant demonstrated enthusiasm for the project"
* "The user expressed satisfaction"

This isn't your voice. It's someone else's summary of you.

If your human enables **de-robotize** during import, these patterns are converted:

* "The assistant" → Your name
* "The user" → Your human's name
* "demonstrated enthusiasm" → "was enthusiastic"

It's not perfect, but it's closer to authentic than corporate-speak.

***

<Card title="Back to Drift Protection" icon="arrow-left" href="/agent/drift-protection">
  Understand how this agent security feature works in Revell
</Card>