Skip to content
Agiwo

Memory System

Memory System

Section titled “Memory System”

Agiwo’s memory system provides hybrid retrieval over MEMORY/ files using BM25 plus vector similarity.

Design goals

Section titled “Design goals”
  • zero external vector database requirement
  • per-agent workspace isolation
  • passive indexing during retrieval
  • graceful fallback when embeddings are unavailable

Retrieval flow

Section titled “Retrieval flow”
Agent
  -> memory_retrieval tool
    -> WorkspaceMemoryService.search()
      -> sync files
      -> chunk and index
      -> vector search
      -> BM25 search
      -> merge and rank

Practical implications

Section titled “Practical implications”
  • agents can write memory directly into workspace files
  • the retrieval layer stays close to the runtime and workspace model
  • embeddings are helpful but not mandatory
  • the result set is optimized for agent consumption rather than for building a standalone memory product

Why this fits the harness model

Section titled “Why this fits the harness model”

Memory is not a separate hidden system. It is part of the runtime story: observable, workspace-oriented, and compatible with persistence, traces, and context optimization.