Source-Filtered Retrieval

Every memory in Aurra carries a source field describing where it came from. Source-filtered retrieval lets you answer questions like "what did my agent learn from Slack this week" or "query only facts captured during chat sessions".

The source field

When you write a memory via POST /agent/memories, source is an optional string that tags the memory's origin.

Conventions Aurra recognizes:

You can also use custom strings (git_commit, calendar, transcript, etc.). Aurra does not enforce a fixed enum.

Filtering on query

Pass sources as an array on POST /agent/query:

curl -X POST https://api.aurra.us/agent/query \
  -H "Authorization: Bearer $AURRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question": "What did engineering decide about pagination?", "sources": ["slack"]}'

Retrieval excludes every memory with a different source. If no memories match the filter (or none match the semantic query within the filtered subset), you get the explicit empty results response (see citations).

The filter is applied at SQL before vector search, so it's cheap and deterministic (not an LLM choice).

Filtering on list

Pass source as a query-param string (not an array) on GET /agent/memories:

curl "https://api.aurra.us/agent/memories?source=slack&limit=50" \
  -H "Authorization: Bearer $AURRA_API_KEY"

Useful for building source-segmented dashboards or running source-specific cleanup jobs.

Patterns

Multi-source agents

If your agent ingests from multiple systems, tag each write with the right source so you can answer "where did this knowledge come from" queries later.

# Slack ingestion
client.memories.add(
    content="Engineering decided to ship cursor pagination instead of offset.",
    source="slack",
    tenant_id="acme-demo",
)

# Notion ingestion
client.memories.add(
    content="Roadmap for Q3 includes an api rate-limiting overhaul.",
    source="notion",
    tenant_id="acme-demo",
)

# Later: credibility query scoped to engineering chat
answer = client.agent.query(
    question="What did engineering decide about pagination?",
    sources=["slack"],
    tenant_id="acme-demo",
)

Trust-level routing

Some sources are high-trust (CRM after a human review) and some are low-trust (auto-extracted chat). You can use source filters to route queries appropriately:

• For surfacing facts in a customer-facing response: restrict to crm + email.
• For internal debugging: allow all sources.
• For compliance reports: restrict to explicit-consent sources only.

Per-integration debugging

To verify your Slack ingestion is working, query with source: slack and confirm counts match expectations. If counts are zero, the ingestion pipeline is broken - not retrieval.

Original input preservation

Every memory also carries a source_citation object with the original input text and capture timestamp. Source filtering trims the search space; source_citation preserves the paper trail. See provenance in the audit endpoint.

Next steps

• Query API - full field reference for sources.
• Memories API - complete source query-param docs.
• Audit API - inspect provenance.original_input for any stored memory.
• BYO-LLM - swap the extraction model per source if needed.