Concepts

Human-in-the-loop memory

Show a person what your agent remembers about them, and let them correct it, forget it or erase it.

Korely turns each memory your agent saves into short facts, such as Sofia / works_at / Acme Corp. Each fact belongs to one end user and has an id. Your product can show a "what do you remember about me" screen and correct or forget a fact with one call.

Terms

The Glossary defines the words used across the docs. These are the ones specific to this page.

Human in the loop: the person a memory is about can see, correct and remove it.
Correct (supersede) a fact: replace a fact with a corrected one. The old fact becomes a closed fact, kept as history.
Forget a fact: close one fact, so reads stop returning it. Forget, delete or erase compares it with the other two.
Source memory: the memory a fact was read from. Each fact names it in source_memory_id.

Quick start

Install pip install korely-memory (Python 3.9 or later) or npm install korely-memory (Node.js 18 or later). Then get a free API key and run export KORELY_API_KEY=kor_live_.... Tested with SDK 0.1.28 (Python) and 0.1.17 (Node.js).

The code saves one memory for customer-4812, waits for extraction, shows the facts, corrects the Acme fact, reads the history and forgets the correction.

import time
from korely_memory import Korely

korely = Korely()  # reads KORELY_API_KEY from the environment
USER = "customer-4812"

mem = korely.add("Sofia works at Acme Corp as a product manager.", user_id=USER)
while korely.get(mem.id).status == "processing":
    time.sleep(2)  # extraction takes about ten seconds

# 1. Show the facts of this end user.
facts = korely.get_facts(user_id=USER)
acme = next((f for f in facts if "Acme" in (f.object or "")), None)
if acme is None:
    raise SystemExit("No Acme fact: see the note below.")
print(acme.subject, acme.predicate, acme.object)

# 2. Correct: Sofia moved to Beacon Labs.
fixed = korely.correct_fact(acme.id, object="Beacon Labs")

# 3. History, closed facts included.
for f in korely.get_facts(user_id=USER, include_invalidated=True):
    print(f.object, f.invalid_at)

# 4. Forget.
print(korely.forget_fact(fixed.id).status)

import { Korely } from 'korely-memory';

const korely = new Korely(); // reads KORELY_API_KEY from the environment
const user_id = 'customer-4812';
const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));

const mem = await korely.add('Sofia works at Acme Corp as a product manager.', { user_id });
while ((await korely.get(mem.id!)).status === 'processing') await sleep(2000);

// 1. Show the facts of this end user.
const facts = await korely.getFacts({ user_id });
const acme = facts.find((f) => f.object?.includes('Acme'));
if (!acme) throw new Error('No Acme fact: see the note below.');
console.log(acme.subject, acme.predicate, acme.object);

// 2. Correct: Sofia moved to Beacon Labs.
const fixed = await korely.correctFact(acme.id!, { object: 'Beacon Labs' });

// 3. History, closed facts included.
for (const f of await korely.getFacts({ user_id, include_invalidated: true }))
  console.log(f.object, f.invalid_at ?? null);

// 4. Forget.
console.log((await korely.forgetFact(fixed.id!)).status);

API=https://api.korely.ai/v1
AUTH="Authorization: Bearer $KORELY_API_KEY"
JSON="Content-Type: application/json"

curl -s -X POST $API/memories -H "$AUTH" -H "$JSON" \
  -d '{"content": "Sofia works at Acme Corp as a product manager.", "user_id": "customer-4812"}'
sleep 15  # wait for extraction

# 1. Show: the id of the Acme fact.
FACT=$(curl -s "$API/facts?user_id=customer-4812" -H "$AUTH" \
  | jq -r '.facts[] | select(.object | contains("Acme")) | .id' | head -n 1)

# 2. Correct: keep the new fact's id.
NEW=$(curl -s -X PATCH $API/facts/$FACT -H "$AUTH" -H "$JSON" \
  -d '{"object": "Beacon Labs"}' | jq -r .id)

# 3. History.
curl -s "$API/facts?user_id=customer-4812&include_invalidated=true" -H "$AUTH" \
  | jq -c '.facts[] | [.object, .invalid_at]'

# 4. Forget.
curl -s -X POST $API/facts/$NEW/forget -H "$AUTH" | jq -r .status

You see lines like these (illustrative):

Sofia works_at Acme Corp
Beacon Labs None
Acme Corp 2026-06-09T14:02:11.418203+00:00
forgotten

Beacon Labs is the current fact: its invalid_at is empty (null in Node.js and JSON). The Acme fact is closed, not deleted. forgotten means Beacon Labs is now closed too.

A memory can finish with status ready and no facts. If there is no Acme fact, save the sentence again, simple and in the present tense, with the same API key and user_id.

Show a person what the agent knows

korely.get_profile(user_id=...) (GET /v1/profile) returns the current facts of one end user, at most 200. by_family groups them by family: the group a predicate belongs to, such as work or health. truncated is true when there are more than 200.

profile = korely.get_profile(user_id="customer-4812")
for family, facts in profile.by_family.items():
    for f in facts:
        print(family, f.id, f.subject, f.predicate, f.object, f.source_memory_id)
print(profile.total, profile.truncated)

For more than 200 facts, page with get_facts (limit up to 200, default 50, and offset). Show source_memory_id so the person can open the sentence behind a fact. Always pass user_id: without it, fact reads cover every end user of the project. Reference: get_profile, get_facts.

Let the person correct a fact

korely.correct_fact(fact_id, object="Beacon Labs") (PATCH /v1/facts/{id}) replaces a fact with a corrected one. Pass at least one of subject, predicate or object. The answer is the new fact, and the next read returns it.

The new fact is current from the moment of the correction.
The old fact is closed at that moment, kept as history and linked to the new one.
When a newer fact contradicts an older one, Korely may close the older one. So a correction can close other facts too: the answer lists them all in invalidated.
The new fact keeps the old source memory, which says Acme Corp. A correction stays only while that memory is neither edited nor deleted: if the memory is edited, the new text is read again and the correction is lost; if it is deleted, the corrected fact is closed too.

Korely can miss a change; a correction fixes it. Only a current fact can be corrected: on a closed one the answer is 409, with current_fact_id to correct instead. A correction asks a language model. If none answers, you get 503 and nothing is written, so retry. Reference: correct_fact.

Let the person forget a fact

korely.forget_fact(fact_id) (POST /v1/facts/{id}/forget) closes the fact. It leaves get_facts, get_profile and the facts part of the context block at once. It stays readable as history with include_invalidated=True.

A forget asks no model and counts no write. Optionally pass at, an ISO date, for when the fact stopped being true. A fact already closed answers already_forgotten; an unknown id answers 404. Reference: forget_fact.

Forgetting a fact leaves the memory that stated it. Search, the memory list and get_context can still return that sentence. If the memory still states another true fact, get_context shows it as plain text, not marked as history, so your agent can still read the old value. To remove it, delete or erase the memory.

A forgotten or corrected value can come back if its memory is edited (PATCH /v1/memories/{id}) or the same sentence is saved again. To avoid it, do not edit that memory or repeat the sentence.

Forget, delete or erase

Action	What it does	What is left
Forget a fact (`forget_fact`)	Closes one fact.	The fact as history, and its memory.
Delete a memory (`DELETE /v1/memories/{id}`)	Hides the memory from every read and closes the facts only it stated. A fact it had closed can become true again.	The row in the database. No API call brings it back.
Erase (`?erase=true` on one memory, or a whole end user)	Removes rows, text and vectors for good.	Nothing, except encrypted backups for up to 35 days.

Erase everything about one person

When a person closes their account in your app, call DELETE /v1/users/{end_user}/memories (SDK: korely.delete_all(user_id=...)). It deletes everything Korely stored for that end user in the key's project, for good, closed facts included. Invented values:

curl -X DELETE https://api.korely.ai/v1/users/customer-4812/memories \
  -H "Authorization: Bearer $KORELY_API_KEY"

# 200 OK
{
  "user_id": "customer-4812",
  "memories_deleted": 218,
  "facts_deleted": 64,
  "erasure": "permanent",
  "audit_id": "aud_91xb"
}

Korely writes one audit record with the counts; audit_id is your receipt.
An end user in several projects needs one call per project, with that project's key.
Memories without a user_id stay. A wrong id answers 200 with zero counts.

Reference: erase an end user.

Use the same controls in the dashboard

The Facts page of the dashboard at agent.korely.ai lists the facts of the selected project, a page at a time. You can filter by end user and family and turn on "Show history". Each current fact has Correct and Forget. A correction there also counts as one write. The Graph page has a slider that shows the facts as they were on a date between the first fact and today.

Details

Reads after a change. Fact reads use no model and see a correction or a forget at once. A copy cached in your agent survives a forget, so read the facts again at the start of each session.
Costs. A correction counts as one write (429 when the month's writes are used); forget and erase count none. Each read counts against your plan's monthly reads, under a rate limit. Read quotas are much larger than write quotas: see pricing.
Permissions. Correct, forget and erase need memories:write; get_profile and get_facts need memories:read. A confined key gets 404 on other end users' facts.
Check ownership. A project key can correct or forget any end user's fact by id: your app must check that it belongs to the person signed in.
Surfaces. REST, the SDKs and the CLI (korely correct-fact, forget-fact, profile, delete-all). The MCP server has korely_correct_fact (close=true forgets a fact) and korely_forget_memory, but no profile or delete-all.
Plans and region. All of this works on every Cloud plan. Stored data sits on EU infrastructure. Each project picks the region of the model that reads its memories, eu or global: see Data governance. For healthcare, legal or support apps, the regulatory review of your application stays yours.