@mono-agent/memory-search

Category

Category: context

Responsibility

Local semantic search over memory. It embeds text with a pluggable provider (Ollama nomic-embed-text by default, OpenAI as a fallback) and stores vectors in a dependency-free JSON Lines index searched by brute-force cosine similarity — fast and simple well under ~50k chunks, with no external vector database or native build. It also gathers indexable chunks from the journal archive and an entity snapshot.

Install / Usage

pnpm --filter @mono-agent/memory-search run build

import { createEmbeddingProvider, createVectorMemoryIndex, gatherMemoryChunks } from "@mono-agent/memory-search";

const embeddings = createEmbeddingProvider({ provider: "ollama", model: "nomic-embed-text" });
const index = createVectorMemoryIndex({ path: "./.mono-agent/memory/index/embeddings.jsonl", embeddings });
await index.rebuild(await gatherMemoryChunks("./.mono-agent/memory", entities));
const hits = await index.search("what did we decide about pricing?");

Public API

• createEmbeddingProvider, OllamaEmbeddingProvider, OpenAIEmbeddingProvider
• createVectorMemoryIndex, VectorMemoryIndex
• gatherMemoryChunks
• MemorySearchError
• EmbeddingProvider, EmbeddingProviderConfig, MemoryChunk, SearchHit, VectorMemoryIndexOptions, EntityLike, MemorySearchErrorCode

Dependency Boundary

This package depends only on the local filesystem and the Fetch API (for the embedding service). It has no other runtime dependencies and no native build. The caller decides when to rebuild the index (typically the nightly consolidation job).

What This Package Does Not Own

It does not own the journal or graph storage, does not extract entities, does not schedule consolidation, and does not expose MCP tools. It only embeds, indexes, and ranks. Hybrid recall (FTS + vector) is orchestrated by @mono-agent/memory-store.

Verification

pnpm --filter @mono-agent/memory-search run build
pnpm --filter @mono-agent/memory-search run typecheck
pnpm --filter @mono-agent/memory-search run test