3rd-Party Memory Integration

Use TIE’s memory as a standalone service from any backend — even if your users never sign in to TIE and your app runs its own LLM. Two endpoints cover the whole loop: write facts into a user’s knowledge graph, and retrieve a prompt-ready context block with the same recall quality as TIE’s own chat agents.

Before You Start

Send your user’s ID in the X-External-User-Id header on every call. The value is case-sensitive and must stay stable — see the header rules.

Step 1 — Write memories

Push facts into the user’s knowledge graph whenever something worth remembering happens in your app — a profile update, an onboarding answer, a meaningful conversation turn.

curl -X POST https://your-tie-host/memories/graph \
  -H "Authorization: Bearer $AUTH_SECRET" \
  -H "X-External-User-Id: crm-user-8841" \
  -H "Content-Type: application/json" \
  -d '{
    "data": "User runs a bakery in Penang and wants to automate invoicing.",
    "type": "text",
    "agent_id": "chatbot",
    "source_description": "CRM profile sync"
  }'

Returns 202 immediately; entity extraction runs in the background. Pass "wait": true to block until ingestion completes (useful in tests). Extraction quality depends on how you phrase the data — write declarative statements, not questions. See Formatting Data for Extraction.

Step 2 — Retrieve context for your LLM

Before each of your own LLM calls, fetch the user’s memory context:

curl -X POST https://your-tie-host/memories/retrieve \
  -H "Authorization: Bearer $AUTH_SECRET" \
  -H "X-External-User-Id: crm-user-8841" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "how can I save time on paperwork?",
    "agent_id": "chatbot",
    "display_name": "Aisha"
  }'

This runs the exact pipeline TIE uses for automatic memory injection in its own chat endpoint — a cached session-context block plus a per-message graph search. The response includes two rendered prompt strings; append them to your system prompt as-is:

context = response["context"]["prompt"]      # "## Known Entities ... ## Known Facts ..."
retrieval = response["retrieval"]["prompt"]  # "## Relevant memories for this message ..."
system_prompt = f"{your_instructions}\n\n{context}\n\n{retrieval}"

Full request/response reference: Retrieve Memory Context. If you only want raw ranked facts to format yourself, use Search Memory instead.

Optional — warm the cache on app open

The context block is cached for 5 minutes per user. To make the first retrieve of a session fast, warm the cache when the user opens your app. Note the warm endpoint takes the derived UUID in its path, not your external ID — in practice it’s simpler to let the first retrieve call warm the cache itself.

Good to know

• One ID, one graph — everywhere. The same external ID always maps to the same user, across all apps sharing the TIE instance. Namespace app-local IDs (myapp:42) if they aren’t globally meaningful.
• Case-sensitivity. Alice and alice are different users. Canonicalize before sending.
• Writes are eventually consistent. A 202 means accepted, not ingested — allow a few seconds before expecting new facts in retrieval, or use "wait": true.