Syntarus developer platform

Syntarus documentation

Connect Syntarus in five minutes: create a key, save a conversation, wait for it to finish processing, then retrieve the small piece of context your agent needs.

THE FIVE-MINUTE PATH

Start with one user and one memory

You do not need to know anything about Qdrant, Neo4j, embeddings, or pipelines. Your agent only needs two habits: save the conversation after a turn, and search before the next answer.

1CaptureSend the conversation or action.
2ProcessSyntarus extracts useful context.
3RetrieveAsk for focused evidence.
4AnswerGive that evidence to your model.

Before you start

  • A Syntarus project and sk_mem_... key
  • A stable user_id per end user
  • Your key stored on your server, never in a browser

What Syntarus handles

  • Durable event processing and status
  • Relevant retrieval across conversations
  • Isolation, deletion, export, and audit controls

HOW IT FITS

Wire memory into any agent

Put the search result in the model prompt. Keep the rest of your agent unchanged.

<memory>
{retrieved_context}
</memory>

Answer using the memory above when it is relevant.
If it is not enough, say what is missing.

On every user turn

  1. Search with the user’s question.
  2. Add the evidence to your model context.
  3. Generate the answer.

After every completed turn

  1. Send user and assistant messages to /memories.
  2. Keep the event ID.
  3. Poll or use a webhook until it succeeds.

SDK

Python: copy, paste, run

Install the SDK, add one conversation, then search the context your agent can use. Keep the API key on your server.

pip install syntarus

import os
from syntarus import MemoryClient

memory = MemoryClient(api_key=os.environ["SYNTARUS_API_KEY"])
event = memory.add(
    user_id="customer_123",
    messages=[{"role": "user", "content": "I prefer vegetarian food."}],
)

memory.wait_for_event(event["event_id"])
results = memory.search(
    user_id="customer_123",
    query="What food does this customer prefer?",
    top_k=5,
)

Tip: Use the returned evidence as a small context block in your model request. Do not send the whole conversation every time.

Any language: use HTTPS

Use the same project-scoped Memory API from any language. Keep your key server-side.

Base URL: https://ai.syntarus.com/syntarus-api/v1

curl -X POST https://ai.syntarus.com/syntarus-api/v1/memories/search \
  -H "Authorization: Bearer $SYNTARUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"user_id":"customer_123","query":"What food do they prefer?","top_k":5}'

POST/memories

Queue a conversation for asynchronous memory processing.

curl -X POST https://ai.syntarus.com/syntarus-api/v1/memories \
  -H "Authorization: Bearer $SYNTARUS_API_KEY" \
  -H "Idempotency-Key: conversation-123" \
  -H "Content-Type: application/json" \
  -d '{"user_id":"customer_123","messages":[{"role":"user","content":"I prefer vegetarian food."}]}'

capture_contract selects how the messages are read. The default, "user_profile_v1", is for a person talking to an assistant: facts are about User. Use "agent_records_v1" for structured records an agent writes about third parties (customers, suppliers, orders, promises, cases). It keeps named parties, dates and changes as written, and does not turn them into facts about the user.

metadata.trust marks how much a message can be relied on. "trusted" is the default for facts the operator or account owner stated. Use "untrusted" for a claim made by someone else, such as a counterparty saying a debt was already paid. Untrusted records are stored and searchable, but they never mark an existing fact as outdated, so an unverified claim cannot overwrite a trusted one.

curl -X POST https://ai.syntarus.com/syntarus-api/v1/memories \
  -H "Authorization: Bearer $SYNTARUS_API_KEY" \
  -H "Idempotency-Key: supplier-claim-7781" \
  -H "Content-Type: application/json" \
  -d '{"user_id":"shop_42","ingestion_mode":"hybrid","capture_contract":"agent_records_v1","source_id":"whatsapp-7781","messages":[{"role":"user","content":"Sharma Traders says they already paid the Rs 40,000 last week."}],"metadata":{"trust":"untrusted"}}'

GET/events/{event_id}

Poll until an asynchronous memory write is succeeded or failed.

POST/memories/search

Retrieve relevant memories for a user and optional agent scope.

{"user_id":"customer_123","query":"What food do they prefer?","top_k":10}

POST/memories/resolve

Retrieve evidence and resolve it into a chronology-aware state card. Use resolution.context as focused agent context and retain results for citations. Requires memories:read and counts toward the retrieval allowance. Your agent remains responsible for decisions and actions.

{"user_id":"customer_123","query":"What is the current support case state?","top_k":10,"max_claims":8,"second_hop_top_k":5}

Using a coding assistant? Read the public MCP setup guide →

POST/bulk/manifests

Estimate a bounded import with start:false, then resubmit with start:true. Each record requires a stable source_id and optional version; imports preserve turn order and run on the lower-priority queue. Poll /bulk/manifests/{manifest_id}, or pause, resume, cancel, and export failed records.

{"start":false,"records":[{"source_id":"ticket-123","source_version":"v1","user_id":"customer_123","messages":[{"role":"user","content":"I prefer email."}]}]}

GET/memories/export

Export a user namespace. Requires the memories:export scope.

DELETE/memories/{memory_id}

Delete one memory. Requires the memories:delete scope.

DELETE/memories?user_id=...

Synchronously delete an entire user namespace. For an auditable cross-store receipt, use a verified deletion job.

POST/platform/deletion-jobs

Queue durable deletion for one user and optional agent namespace. Poll the returned identifier through /events/{event_id}; the completed result contains a per-store verification receipt. Requires memories:delete.

POST/platform/tokens

Issue a short-lived token bound to one user_id and optional agent_id. The requested scopes must be a subset of the parent key and project policy. Requires tokens:write.

GET/platform/audit

Read project-scoped, metadata-only audit events. Memory content and credentials are excluded. Requires audit:read.

WEBHOOKS/webhooks

Create a signed delivery endpoint for memory.ingest.succeeded and data.deletion.succeeded. Requires webhooks:write.

PRODUCTION CHECKLIST

Ship safely

01Use least privilege — start with memories:read and memories:write.
02Keep users isolated — use stable IDs and one project per environment.
03Make writes reliable — send an idempotency key and handle retries.
04Respect user controls — implement export and deletion for user requests.
05Measure the loop — track write success, search latency, and answer quality.

Ready to try it?

Start in the console

Create a project key and test your first memory without writing infrastructure.

Open developer console →

Going enterprise

Use scoped keys, signed webhooks, audit metadata, deletion jobs, and customer-cloud deployment when your workload requires it.