recall

recall

Retrieve data from semantic memory. A recall step performs a similarity search against a vector store collection, returning the most relevant entries for a given query. This is a governed effect: the runtime checks memory permissions, records the retrieval in the behavioral ledger, and mediates the read through the governance interpreter.

When to use

Use recall when you need to:

• Find relevant documents or facts for an LLM prompt (RAG pattern)
• Search conversation history for context
• Look up similar past cases or examples
• Retrieve knowledge base entries matching a query

Use compute for in-memory lookups on data already in scope. Use action ... db for structured SQL queries. Use recall specifically for semantic/vector similarity search that pairs with remember.

Syntax

recall <name>
  query: <expression>
  collection: "<collection_name>"
  limit: <number>
  threshold: <number>
  filter: { <key>: <value>, ... }
  namespace: "<namespace>"

Configuration

Config	Required	Description
query	Yes	The search query. Embedded at runtime and compared against stored vectors.
collection	Yes	Which collection to search. Must match a collection used in remember.
limit	No	Maximum number of results to return. Default: 5.
threshold	No	Minimum similarity score (0.0 to 1.0). Results below this score are excluded. Default: 0.0 (no threshold).
filter	No	Metadata filter. Only entries matching all specified key-value pairs are searched.
namespace	No	Restrict search to a specific namespace partition.

Output shape

A recall step produces an object with:

Field	Type	Description
results	list	Array of matching entries, each with content, metadata, and score
count	number	Number of results returned

Each result entry:

Field	Type	Description
content	text	The stored content
metadata	map	The metadata attached during remember
score	number	Similarity score (0.0 to 1.0, higher is more similar)

Examples

RAG pattern: retrieve context for LLM

machine question_answerer
  accepts
    question as text, is required

  responds with
    answer as text
    sources as list

  implements
    recall find_relevant_docs
      query: input.question
      collection: "knowledge_base"
      limit: 5
      threshold: 0.7

    compute build_context
      let docs = steps.find_relevant_docs.results
      {
        context: docs.map(d => d.content).join("\n\n---\n\n"),
        source_urls: docs.map(d => d.metadata.source)
      }

    ask answer_with_context, using: "anthropic:claude-sonnet-4-6"
      with role "You are a knowledgeable assistant. Answer based only on the provided context. If the context doesn't contain the answer, say so."
      with task "Question: ${input.question}\n\nContext:\n${build_context.context}"
      returns
        answer as text
      assuming
        answer: "Based on the provided context..."

    compute format_output
      {
        answer: steps.answer_with_context.answer,
        sources: build_context.source_urls
      }

Filtered recall with namespace

recall find_recent_feedback
  query: "product quality issues"
  collection: "customer_feedback"
  filter: {category: "complaint", urgency: "high"}
  namespace: "2026-q1"
  limit: 10
  threshold: 0.6

Recall with conditional follow-up

machine research_assistant
  accepts
    question as text, is required

  implements
    recall check_existing
      query: input.question
      collection: "research_cache"
      limit: 3
      threshold: 0.85

    decide use_cache_or_research
      when steps.check_existing.count is greater than 0
        compute cached_answer
          {
            answer: steps.check_existing.results[0].content,
            source: "cache",
            confidence: steps.check_existing.results[0].score
          }
      otherwise
        ask fresh_research, using: "anthropic:claude-sonnet-4-6"
          with task "Research this question thoroughly: ${input.question}"
          returns
            answer as text
          assuming
            answer: "Research findings..."

Governance

Every recall step is governed:

1. Permission check: the machine must have memory capability
2. Read mediation: the governance interpreter mediates the retrieval
3. Behavioral ledger: the query, collection, filter criteria, and result count are recorded
4. Redaction: if the records > redaction section specifies PII redaction, recalled content may be redacted before being passed to downstream steps

In test mode, recall steps return empty results by default. Use assuming in verifies tests to mock recall results for deterministic testing.

Translations

Language	Keyword
English	recall
Spanish	recuerda
French	rappelle
German	erinnert
Japanese	思い出す
Chinese	回忆
Korean	회상

See also

• remember - Store data in semantic memory
• ask … using - Often follows recall in a RAG pattern
• implements - The section where recall steps live
• ensures - Permission declarations for memory capability