@xfabric/memory

Semantic memory search for AI agents. SQLite-backed with vector + FTS hybrid search.

Architecture

graph TB
    subgraph Input
        MD[Markdown Files]
        SESS[Session Transcripts]
    end

    subgraph Core
        MM[MemoryManager]
        CHUNK[Chunker]
        EMB[Embedding Provider]
    end

    subgraph Providers
        OAI[OpenAI]
        GEM[Gemini]
        LOC[Local/llama.cpp]
    end

    subgraph Storage
        DB[(SQLite)]
        FTS[FTS5]
        VEC[sqlite-vec]
    end

    subgraph Search
        VS[Vector Search]
        KS[Keyword Search]
        HYB[Hybrid Merge]
    end

    MD --> MM
    SESS --> MM
    MM --> CHUNK
    CHUNK --> EMB
    EMB --> OAI
    EMB --> GEM
    EMB --> LOC
    EMB --> DB
    DB --> FTS
    DB --> VEC
    FTS --> KS
    VEC --> VS
    VS --> HYB
    KS --> HYB
    HYB --> MM

Install

pnpm add @xfabric/memory

Usage

import { MemoryManager } from "@xfabric/memory/full";

const memory = await MemoryManager.create({
  agentId: "my-agent",
  workspaceDir: "/path/to/workspace",
  provider: "openai", // or "gemini", "local"
});

// Search
const results = await memory.search("authentication flow");

// Write memory
await memory.writeMemory("## Today\n- Fixed auth bug");

// Close
await memory.close();

Features

• Hybrid search: Vector similarity + FTS5 keyword matching
• Multiple providers: OpenAI, Gemini, local (llama.cpp)
• Batch API: Async embedding for large workspaces
• Auto-sync: File watcher with debouncing
• Session indexing: Index conversation transcripts
• Memory flush: LLM-powered extraction from sessions

License

MIT