Recall Flow

Unified Recall Retrieval

v1 ✅ Implemented: memory_recall(query) four-source unified recall, returns results[] (distinguished by memory_type).

v1 Implemented

Capability	Description
memory_recall	Four sources: semantic / conversation / limbic / autobiographical; Top N; conversation returns snippet
sessions_search	Session message search; returns snippet (not full content)
memory_remember	Semantic memory CRUD during conversation
Resident memory	System prompt injection (not recall)

Four Sources

memory_type	Content type
semantic	Facts, preferences, experiences, etc.
session	Historical conversation snippets
limbic	Emotional memory
autobiographical	Narrative title + content snippet

Resident Memory

Separate from recall: pinned semantic memories plus most-referenced entries are always injected into the system prompt. LLM cites memory ID markers in replies; reference counts sync via nightly cron.

Relationship to Self Layer

Self layer (six blocks) is not retrieved via memory_recall—it is always in the system prompt.

Layer	Injection	Reason
Self layer	Always in system prompt	Small, fixed, every conversation needs “who I am”
Memory layer	memory_recall on demand	Large, dynamic, search when needed

Future Direction

Unified recall extending to resource layer (tools, skills, assets) with preset weighting is tracked in Issue #47. Not current runtime behavior.

Naming Rationale

In cognitive psychology, Recall is actively pulling stored information from memory—retrieving what you already stored. Even when scope extends to resources, the action’s essence is unchanged.