epistemic_recall

Query the memory store using natural language. Returns ranked results with confidence scores.

QUERY

epistemic_recall

Description

Performs hybrid search combining vector similarity (cosine) and keyword matching. Results are ranked by relevance and filtered by tier visibility rules. Access timestamps are refreshed on recall (decay reset).

Parameters

Name	Type	Required	Description
`query`	`string`	Yes	Natural language query (e.g., "What food does the user like?").
`limit`	`number`	No	Maximum results to return. Default: `10`.
`minConfidence`	`number`	No	Minimum confidence threshold. Default: `0.0`.
`tier`	`string`	No	Filter by tier: `FACT`, `WORKING`, `CANDIDATE`, `QUARANTINE`, `CHALLENGED`.
`kind`	`string`	No	Filter by claim kind: `identity`, `preference`, `behavioral`, etc.
`subject`	`string`	No	Filter by subject (e.g., `"user"`, `"Lan"`).

Example

Request

{
  "tool": "epistemic_recall",
  "input": {
    "query": "user food preferences",
    "limit": 5
  }
}

Response

{
  "results": [
    {
      "id": "mem_f8a2b1c3",
      "claim": "Tôi thích ăn phở bò Hà Nội",
      "subject": "user",
      "confidence": 0.741,
      "tier": "WORKING",
      "kind": "preference",
      "relevance": 0.923,
      "storedAt": "2026-01-15T10:30:00Z"
    }
  ],
  "total": 1,
  "searchType": "hybrid"
}

Notes

Each recall refreshes lastAccessed, resetting the decay timer
Channel isolation rules apply — non-owner users only see memories from their channel
QUARANTINE tier memories are excluded by default unless explicitly filtered