Migrating From Mem0 / Letta / Zep to Ujex Recall

Migration between agent memory layers is doable but not magical. The shapes differ; the transform isn't identity. This post is the recipe for moving to Ujex Recall.

Prerequisites

• Source: Mem0 / Letta / Zep — production data
• Target: Ujex project (Firebase) with Recall enabled
• Your agent's namespace plan (e.g. user_facts/, conversations/<date>/, episodes/)

Step 1 — Export

From Mem0

from mem0 import Memory
m = Memory()
all_memories = m.get_all(user_id='your-user')
# Returns list of {id, memory, created_at, updated_at, metadata}
import json
with open('mem0-export.json', 'w') as f:
    json.dump(all_memories, f)

From Letta

Letta's archival memory exports via the SDK:

from letta import create_client
client = create_client()
agent = client.get_agent(agent_name='my-agent')
archival = client.get_archival_memory(agent_id=agent.id)
import json
with open('letta-export.json', 'w') as f:
    json.dump([{'text': p.text, 'metadata': p.metadata} for p in archival], f)

From Zep

Zep has a documented export API. GET /api/v2/memories?session_id=... returns JSON.

Step 2 — Transform to Markdown

import json, yaml, hashlib
with open('mem0-export.json') as f:
    rows = json.load(f)

for row in rows:
    name = row.get('id') or hashlib.sha256(row['memory'].encode()).hexdigest()[:12]
    fm = {
        'type': 'memory',
        'created': row.get('created_at'),
        'updated': row.get('updated_at'),
        'tags': row.get('metadata', {}).get('tags', []),
    }
    md = '---\n' + yaml.dump(fm) + '---\n\n' + row['memory']

    with open(f'export/{name}.md', 'w') as out:
        out.write(md)

Step 3 — Upload to your bucket

gcloud storage cp -r export/ gs://your-ujex-bucket/agents/your-agent/memories/

Recall's index updates within a minute (Cloud Function trigger on bucket write).

Step 4 — Re-embed (if you want vector search)

Vertex AI embeddings differ from OpenAI / Cohere — the vectors don't transfer. Trigger re-embed:

from ujex_recall import RecallStore
store = RecallStore(api_key=os.environ['UJEX_API_KEY'], agent_id='your-agent')
store.reindex(namespace='memories')  # walks the bucket prefix, embeds, writes Firestore index

Namespace mapping

What to keep vs drop

Keep: text content, metadata that's still meaningful (tags, type, created/updated), source IDs in frontmatter for back-traceability.

Drop: source-specific embeddings (re-compute), TTL fields (handle in Recall's lifecycle), proprietary fields with no equivalent.

Validation

After ingest:

from ujex_recall import RecallStore
store = RecallStore(api_key=...)
results = store.search(('memories',), query='something I know is in there', limit=5)
assert len(results) > 0

Rollback

Bucket is the source of truth. gcloud storage rm -r gs://your-ujex-bucket/agents/your-agent/memories/ deletes the migration. Your old store is untouched.