Migrate from Mem0

If you built on Mem0, the Korely API will feel familiar. The verbs are the same: add a memory, search memories, list them, delete them. The scoping vocabulary is the same: user_id identifies your end user, agent_id identifies your application, run_id identifies a session. Switching is mechanical: map the calls, replay the corpus through the batch endpoint, verify with a search.

Terms

The Glossary defines the words used across the docs. This page uses memory, fact, end user, agent_id, run_id, metadata, batch, valid_from, extraction, events and context block as defined there.

The Mem0 calls on this page are the MemoryClient of the mem0ai Python package, version 2.x (checked on 2.2.1).

Concept mapping

Mem0 and Korely share the same scoping vocabulary. The table below maps each Mem0 concept to its Korely equivalent so you know what to keep, what to rename, and what is new.

Call mapping

Every core Mem0 call has a Korely equivalent, in the Python SDK and over REST. The full REST contract is in the API reference.

The scoping parameters keep their names and meaning: Mem0 reads take them inside filters, Korely takes them as plain arguments. user_id is your end user's identifier, free-form, and end users are unlimited on every tier. agent_id namespaces your application. run_id sub-scopes one session. Filters are additive.

In code, the switch looks like this:

# before
from mem0 import MemoryClient
m = MemoryClient(api_key="...")
m.add("Prefers invoices as PDF", user_id="customer-4812")

# after
from korely_memory import Korely
korely = Korely(api_key="kor_live_...")
korely.add("Prefers invoices as PDF", user_id="customer-4812")

results = korely.search("invoice preferences", user_id="customer-4812")

Python and Node.js today.

Gotchas

The surface is familiar, but a handful of parameter names and behaviours differ. Check these before running your migration.

Step 1: export from Mem0

Export with Mem0's list call, get_all, page by page. Mem0 requires a filters argument, so list your own end users and read each one. Save every page's results in one file as {"results": [...]}: each item carries the text under memory, plus user_id, metadata and created_at, which map onto the Korely batch payload almost field for field.

import json
from mem0 import MemoryClient

m = MemoryClient(api_key="m0-...")
user_ids = ["customer-4812", "customer-5113"]  # your own user list

export = {"results": []}
for uid in user_ids:
    page = 1
    while True:
        r = m.get_all(filters={"user_id": uid}, page=page, page_size=100)
        export["results"] += r["results"]
        if not r.get("next"):
            break
        page += 1

json.dump(export, open("mem0_export.json", "w"))
print(len(export["results"]), "memories exported")

One thing to know before you import. Mem0 keeps the statements it extracted (for example Prefers invoices as PDF), not the conversation they came from. Each statement becomes one Korely memory, and Korely extracts its facts from that short text. If you still hold the original conversations, importing those gives Korely more to work with.

Step 2: import with the batch endpoint

POST /v1/batch accepts up to 500 memory objects per request, same shape as POST /v1/memories, processed asynchronously. You get a job id back and poll it until it completes. Each imported item runs the full write pipeline: embeddings, entity extraction, typed-fact extraction with contradiction checking and bi-temporal validity. The model work is included in your plan: a memory counts one write per 6,000 characters of its text.

import json, time, requests

export = json.load(open("mem0_export.json"))

def to_korely(item):
    memory = {
        "content": item["memory"],
        "user_id": item.get("user_id"),
        "metadata": {**(item.get("metadata") or {}), "source": "mem0_export"},
    }
    if item.get("created_at"):
        # when it was true: the facts take it as valid_from, not the import day
        memory["timestamp"] = item["created_at"]
    return memory

memories = sorted((to_korely(i) for i in export["results"]), key=lambda m: m.get("timestamp", ""))

def send(chunk):
    while True:
        r = requests.post(
            "https://api.korely.ai/v1/batch",
            headers={"Authorization": "Bearer kor_live_..."},
            json={"memories": chunk},
        )
        if r.status_code != 429:
            r.raise_for_status()
            return r.json()
        code = r.json().get("code")
        if code == "rate_limit_exceeded":    # wait as long as the server says
            time.sleep(int(r.headers.get("Retry-After", "60")))
        else:                                # quota_exceeded or agent_cap_exceeded: stop
            raise SystemExit(r.json()["message"])

# POST /v1/batch takes up to 500 memories per request
for i in range(0, len(memories), 500):
    print(send(memories[i : i + 500]))  # {"id": "job_4e1aa0", "status": "pending", "received": 500}

Poll the job until it reports completed:

curl https://api.korely.ai/v1/batch/job_4e1aa0 \
  -H "Authorization: Bearer kor_live_..."

# 200 OK
{"id": "job_4e1aa0", "status": "completed", "received": 500, "imported": 500, "failed": 0, "errors": []}

Step 3: verify with search

Pick a query you know your old corpus can answer and run it scoped to one end user. Memory search is semantic vector retrieval, cosine similarity over embeddings. The only model work on the read path is embedding your question, which runs on Korely's own servers. No generative model composes the output. Your agent's own model does the reasoning over the results. For recall assembled from the typed-fact layer rather than raw snippets, reach for GET /v1/context, see below.

curl -X POST https://api.korely.ai/v1/memories/search \
  -H "Authorization: Bearer kor_live_..." \
  -H "Content-Type: application/json" \
  -d '{"query": "invoice preferences", "user_id": "customer-4812", "limit": 5}'

# 200 OK
{
  "results": [
    {"id": "mem_3fa1c9", "score": 0.91,
     "snippet": "Prefers invoices as PDF, replies fastest before 10am CET.",
     "user_id": "customer-4812", "agent_id": null,
     "metadata": {"source": "mem0_export"}}
  ]
}

Extraction runs after the batch job has stored the memories. Once GET /v1/events reports processing: 0, a GET /v1/facts call against a migrated entity confirms the typed layer is populated.

What your agents gain

A typed knowledge graph

Every imported memory goes through entity extraction. People, companies, products, places, and concepts become nodes; relations between them become typed edges, drawn from canonical predicates grouped into families (preferences, people, places, work, ownership, health, financial, events, identity, and an other bucket for predicates outside the named groups). The graph is queryable: GET /v1/facts?entity=... returns every typed fact touching an entity, its neighbourhood in the graph, so you can walk from any subject to what it is connected to. These reads are deterministic SQL with zero AI calls and no embeddings.

# list all memories for a user (SDK)
memories = korely.get_all(user_id="customer-4812")

# or over REST
# GET /v1/memories?user_id=customer-4812&limit=20
# check the graph layer with GET /v1/facts?user_id=customer-4812
# {"facts": [{"subject": "customer-4812", "predicate": "prefers", "predicate_raw": "prefers",
#             "object": "PDF invoices", "predicate_family": "preferences", "invalid_at": null, ...}], "total": 1}

See the graph for a full explanation of how entities connect memories and how to traverse those connections over the REST API.

Bi-temporal facts

Each memory also yields typed (subject, predicate, object) triples, and every triple carries valid_from and invalid_at. When new information contradicts an existing fact, a two-stage contradiction check at write time invalidates the old fact and keeps it as history. Default reads return only what is true now; pass include_invalidated=true for the full supersede chain, or as_of on GET /v1/facts for a point-in-time view. Facts reads are deterministic SQL, with no model call, and return flat JSON, {"facts": [...], "total": N}, with each fact's phrase as extracted preserved in predicate_raw alongside the predicate, the relation that phrase resolved to in the end user's vocabulary (usually the same words in snake_case). The temporal engine runs on every tier, and GET /v1/facts / korely.get_facts() work on every plan, free hobby tier included. The full model is in temporal facts.

Memory your end users can see

Every fact is a readable triple, so your product can show a user what the agent knows (get_profile) and let them fix it: correct_fact supersedes a wrong fact, forget_fact closes one, and delete_all erases everything about the user, for good, with an audit record. The dashboard has Correct and Forget on every current fact. If agents carry long-term memory about people, the people get the delete button. Read human in the loop for the full model.

Underneath all three: reads are retrieval, not generation. The intelligence runs once, at write time, which is why read quotas are a hundred times or more the write quotas. Quotas for each plan are on the pricing page.

Next steps

• Quickstart: wire Korely into your stack and run your first search in five minutes.
• API reference: the complete REST contract behind every call in the mapping table.
• Architecture: where the write-time intelligence runs and why the read path stays deterministic.
• Questions about a large corpus or an unusual export shape? Email [email protected]. We read every message.