75.0% LoCoMo — outperforms Mem0 and GPT-based systems

Add memory to your AI
in 2 API calls

Working memory in under 5 minutes. No infra required.

Stop rebuilding memory logic. CLAIV gives your chatbot, agent, or app persistent, structured memory across conversations and documents.

Works withOpenAIClaudeLangChainLlamaIndex
$ npm install @claiv/memoryTry it now
TypeScript · quickstart.ts
Ingest — store a memoryimport { CLAIV } from '@claiv/memory'

const claiv = new CLAIV({ apiKey })

await claiv.ingest(({
  user_id:  'user-123',
  conversation_id: 'conv-456',
  type:  'message',
  role:  'user',
  content: 'I run a fitness business'
});Recall — inject into promptconst memory = await claiv.recall(({
  user_id:  'user-123',
  conversation_id: 'conv-456',
  query:  'What does the user do?'
});

// memory.llm_context.text →
// "User runs a fitness business"

systemPrompt += memory.llm_context.text
That's the entire integration. No schemas. No pipelines. No memory logic to write.

Stop building this yourself

Every team rebuilds the same memory layer. Most ship something fragile. Here's what you're replacing.

What you're doing today
// Memory logic you wrote — and now maintain
storeChatHistory(messages)
buildVectorSearch(embeddings)
rerankResults(chunks)
handleContradictions(facts)
manageTokenLimits(context)
pruneOldContext(history)
handleGDPRDeletion(userId)

~8 systems · ~weeks of engineering · ongoing maintenance

What you do with CLAIV
// Two calls. That's the entire layer.
claiv.ingest({ user_id, content })
claiv.recall({ user_id, query })

// That's it.
// Extraction → automatic
// Contradiction handling → automatic
// Token budgeting → automatic
// GDPR receipts → automatic

Working in < 5 minutes · no infra to manage

The problem

Your AI doesn't have memory.
It has a token limit.

This is not a model problem. It's a memory problem.

Users repeat themselves every session

"I already told you my name, my preferences, my project details." Without persistent memory, every conversation resets.

Context gets truncated silently

Stuffing conversation history into prompts doesn't scale. Long threads get cut off mid-thought. Critical facts disappear without warning.

RAG returns similar text, not the right fact

Embedding-based retrieval finds "similar" text, not the specific fact you need. Contradictions pass through undetected.

No way to actually delete user data properly

GDPR requires provable deletion. "We asked the model to forget" isn't compliance. You need timestamped, auditable proof.

The solution

Three API calls. That's it.

No schemas. No pipelines. No memory logic to build.

01

Ingest

POST /v6/ingest

Store messages → structured facts extracted automatically. Contradictions resolved. Timeline tracked. Evidence spans preserved.

02

Recall

POST /v6/recall

Get ranked context → ready for your system prompt. Fact retrieval ranked within your token budget. One field to inject.

03

Forget

POST /v6/forget

Delete user data → with audit-proof receipt. GDPR-compliant deletion with timestamped documentation of exactly what was removed.

Under the hood

What happens when you ingest

Deterministic. Testable. Production-safe.

1

You send messages

Any conversation turn — user or assistant. You send it to /v6/ingest via API.

2

Facts extracted async

CLAIV parses structured facts from the text. Contradictions are resolved. Temporal links are built.

3

Versioned and indexed

Facts are timestamped, versioned, and tiered. High-importance facts always appear in recall.

4

Recall within token budget

You specify a token limit. CLAIV ranks, fits, and returns a ready-to-inject context string.

Differentiation

Not a vector database
with a memory label

Vector search finds similar text. CLAIV stores structured facts.

Vector DB
CLAIV Memory
Retrieval method
Similarity search
Structured fact lookup
Contradictions
Not handled
Resolved automatically
Temporal awareness
None
Full timeline tracking
Token budget control
Manual chunking required
Built-in ranked fitting
GDPR deletion proof
No audit receipt
Timestamped audit receipt
Fact provenance
No source tracking
Evidence spans per fact

Document Memory

RAG that actually works

Upload documents once. CLAIV parses structure, indexes content, and injects the right sections into memory automatically — alongside conversation facts.

Section-aware retrieval

Not chunk spam. CLAIV understands document structure and surfaces the right section in context.

Works alongside conversation memory

Document facts and conversation facts are recalled together in a single /v6/recall call. One injection.

No pipeline to maintain

Upload via POST /v6/documents. Chunking, embedding, and indexing handled automatically.

POST /v6/documents
curl -X POST https://api.claiv.io/v6/documents \
  -H "Authorization: Bearer <API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "user_42",
    "project_id": "proj_abc",
    "document_name": "q4-roadmap.txt",
    "content": "Q4 roadmap: ship RAG..."
  }'

200 OK

{
  "document_id": "doc_7fx9qr",
  "sections": [{ "title": "Overview" }],
  "spans_created": 14,
  "status": "processing"
}
Independent Benchmark

We outperform typical memory systems

Tested across 10 dialogue datasets, 1,540 questions covering recall, temporal reasoning, and generation. Outperforms Mem0 and GPT-based approaches on their own published results.

Scores by category

Open-Domain GenerationFree-form answers synthesised across full conversational history
79.7%
Temporal ReasoningTime-ordering and duration questions derived from event sequences
74.2%
Single-Hop RecallSingle fact retrieval from one location in conversation history
68.8%
Multi-Hop ReasoningEvidence chaining across multiple non-adjacent conversation turns
55.2%

Overall score

75.0%

vs competitors

CLAIV V6.475%
Mem066.88%
Rules GPT52.9%

All 10 LoCoMo dialogue sets · GPT-4o mini LLM judge · Best published results per provider

Use cases

If your product has a chat interface,CLAIV gives it memory.

AI chatbots with memory

Users never repeat themselves. Preferences, context, and history persist across every session.

AI agents with workflows

Agents remember task state, user goals, and prior actions. Long-running workflows stay coherent.

Internal copilots

Retain institutional knowledge, team conventions, and workflow preferences across your organisation.

Customer support systems

Know the customer's history before they speak. Resolve faster. Escalate with full context.

Pricing

Pay for memory usage.
Nothing else.

No per-user fees. No seat licences. You pay for ingests — messages stored into memory.

Typical usage example

~30 messages/month

~$0.12/month

that's it — no hidden fees

Free

$0

200 ingests/mo

Starter

$29/mo

5K ingests/mo

Growth

$99/mo

20K ingests/mo

Scale

$299/mo

75K ingests/mo

Stop rebuilding memory.

Free tier included. No credit card required.