RecallForge

Every modality, one search. Local first.

Standard RAG only works on text. Drop a PDF with charts, a photo of a whiteboard, a video recording, or a transcript-backed audio note — and your AI agent goes blind. RecallForge gives agents eyes and ears over your local filesystem. Text, images, documents, video, and audio transcripts all live in one unified search space, and nothing ever leaves your machine.

What this enables

You: "What did the whiteboard look like in our last meeting?"

Claude: (Searches your local ~/Documents, finds a photo of a whiteboard from an iPhone, reads the handwriting via Qwen3-VL, and surfaces the image with context.)

You: "Find the architecture diagram from that PDF I downloaded last week."

Claude: (Indexes the PDF, matches your query against extracted text and embedded figures, returns the relevant page.)

You: (Drops an image of a circuit board) "Find my notes related to this."

Claude: (Reverse image-to-text search across your indexed notes. Returns matching documents.)

One query. Any modality. All local.

What makes RecallForge different

Capability	RecallForge	Chroma	Mem0	Qdrant	Weaviate
Cross-modal search	✅ Native	✅ OpenCLIP	❌ Text only	❌	✅ CLIP modules
Video support [Beta]	✅	❌	❌	❌	❌
Audio transcript ingest	✅	❌	❌	❌	❌
Document ingest (PDF/DOCX/PPTX)	✅	❌	❌	❌	❌
Built-in reranking	✅ Multimodal	❌	❌	✅ ColBERT	✅ Modules
MCP-native	✅ 26 tools	❌	❌	❌	❌
100% local	✅	✅	⚠️ Cloud default	✅	✅ Docker
Apple Silicon optimized	✅ MLX 4-bit	❌	❌	❌	❌
Cloud option	❌	✅	✅	✅	✅
JS/TS SDK	❌	✅	✅	✅	✅

Use RecallForge when: You need multimodal memory for AI agents that runs entirely on your machine, especially on Apple Silicon. One search across text, images, documents, video, and transcript-backed audio.

Use something else when: You need cloud hosting, massive scale (millions+ vectors), or a JS/TS-first ecosystem.

Performance

5 modalities (text, images, documents, video, transcript-backed audio) unified in a single MLX-optimized local vector space. Sub-60ms search latency in embed mode. Under 400MB resident memory.

Pipeline ablation (Mac mini M4 16GB, MLX 4-bit)

Each stage of the pipeline improves retrieval quality. The reranker is the quality peak.

Stage	R@1	R@5	R@10	MRR	p50
Vector-only	65.2%	65.2%	67.4%	67.3%	20ms
BM25-only	57.6%	57.6%	93.5%	64.4%	17ms
Vector + BM25 (RRF)	69.6%	88.0%	90.2%	77.5%	100ms
+ Reranker (hybrid mode)	85.9%	92.4%	97.8%	89.2%	3.8s

The reranker delivers +20.7% R@1 over RRF fusion and pushes R@10 to 97.8%. Embed mode gives you 20ms searches for speed-sensitive workloads. Hybrid mode gives you 85.9% R@1 when quality matters.

Benchmark categories: text_only (30 queries), image_only (30 queries), long_query (12 queries), typo_query (20 queries). See benchmarks/results/pipeline_ablation_modality_results.json for full breakdown.

For release validation, use benchmarks/cross_modal_ablation.py. It checkpoints JSON output as it runs, so long MLX benchmark sessions still leave behind a partial artifact if interrupted. The UAT video corpus now uses compact episodic fixtures with searchable transcript sidecars and related artifact metadata, so video queries exercise meeting, screen-recording, walkthrough, field, and recipe-style memories. To turn a benchmark artifact into a ranked fix list, run benchmarks/cross_modal_diagnostics.py; the current report is in docs/research/cross-modal-diagnostics.md.

Latency & resource usage

Metric	MLX 4-bit	PyTorch fp16
Warm search p50 (embed)	53ms	599ms
Warm search p95 (embed)	55ms	—
Cold start	7.6s	~20s
Peak RSS (embed)	329MB*	~4GB
Peak RSS (hybrid)	~1.5GB*	~5GB
Text indexing	5.0 docs/sec	—

*MLX maps model weights lazily via memory-mapped files. RSS reflects resident pages, not full model size (~1.7GB embedder + ~1.7GB reranker on disk). Actual memory pressure is low.

COCO 1K retrieval (raw embeddings, no pipeline)

For transparency: raw embedding quality on the standard COCO benchmark (1,000 images, no BM25/reranking/expansion). These numbers reflect the Qwen3-VL-2B embedder alone, not the full pipeline.

Direction	R@1	R@5	R@10
Text → Image	24.5%	42.3%	49.9%
Image → Text	34.3%	42.0%	44.1%

Qwen3-VL is a generative VLM, not a contrastive model like CLIP. The pipeline ablation above shows how BM25 fusion and reranking compensate for this.

Installation

pip install recallforge[mlx]       # Apple Silicon (recommended, 4-bit quantization)
pip install "recallforge[mlx,server]"  # Apple Silicon + HTTP/SSE server
pip install recallforge[cuda]      # NVIDIA GPU
pip install recallforge[torch]     # CPU / other PyTorch targets
pip install recallforge[docs]      # add richer PDF extraction (optional)

Note: pip install recallforge installs the core without a backend. You need at least one of [mlx], [cuda], or [torch] to run inference. Add [server] only when you want HTTP/SSE transport (recallforge serve --http).

From source:

git clone https://github.com/brianmeyer/recallforge.git
cd recallforge
pip install -e ".[mlx]"

Requirements

• Python 3.12 or 3.13 required (3.14 not yet supported, pending pyarrow wheel)
• Disk: ~2-5GB free for model downloads on first run
• RAM (MLX 4-bit): ~1.7GB (embed) to ~3.4GB (hybrid)
• ffmpeg recommended for video indexing/search
• Audio indexing is transcript-first: add a .srt, .vtt, .txt, or .transcript.json sidecar next to the audio file
• First run downloads models automatically and may take a few minutes

MCP Server (primary use)

RecallForge is designed as a Model Context Protocol server for AI agents. Configure in Claude Desktop (or any MCP-compatible agent host):

{
  "mcpServers": {
    "recallforge": {
      "command": "recallforge",
      "args": ["serve", "--mode", "hybrid"]
    }
  }
}

Run manually (stdio):

recallforge serve --mode embed --backend mlx --quantize 4bit

Run over HTTP/SSE:

recallforge serve --http --host 127.0.0.1 --port 7433 --mode embed

RecallForge now exposes 26 MCP tools across search, ingest, memory graph navigation, collection admin, and runtime config. HTTP/SSE mode also exposes /health, /sse, and /messages/. Long-running tools emit MCP notifications/progress when the client supplies a request _meta.progressToken, so compatible HTTP/SSE clients can show live progress for ingest, search, batch, memory writes, and FTS rebuilds.

See docs/mcp-tools.md for the full tool reference.

Search modes

Mode	Models loaded	Memory (MLX 4-bit)	Quality	Best for
embed	Embedder	~1.7GB	Good	Memory-constrained, fast searches
hybrid	Embedder + Reranker	~3.4GB	Best	Maximum retrieval quality

Video [Beta] note: Video support requires ffmpeg. The torch backend video path has a known upstream issue (see QwenLM/Qwen3.5#58).

How it works

RecallForge encodes text, images, video frames, documents, conversation turns, and audio transcripts into the same 2048-dimensional vector space using Qwen3-VL. It also extracts lightweight entity and relation metadata so agents can navigate from one memory to other memories that mention the same people, projects, tickets, URLs, and organizations. Reindexes for documents, video, audio, and conversations are staged as hidden batches first, then promoted together so agents keep seeing the previous complete memory until the replacement is ready. This means "find notes about this diagram" works whether the diagram is text, an image, a conversation thread, or a frame from a video. A 3-stage pipeline handles the rest:

graph TD
    subgraph Local Filesystem
        Docs[📄 Documents]
        Imgs[🖼️ Images]
        Vids[🎬 Video]
        Aud[🎙️ Audio + Transcript]
        Conv[Conversation Turns]
    end

    subgraph RecallForge Ingest
        Docs --> TxtExt[Text Extractor]
        Imgs --> VLM[Qwen3-VL Encoder]
        Vids --> Frame[Frame & Audio Extractor]
        Aud --> TxtExt
        Conv --> TxtExt
        Frame --> VLM
        TxtExt --> VLM
    end

    subgraph LanceDB Storage
        VLM -->|2048-dim Vectors| VecDB[(Vector Space)]
        TxtExt -->|Text/Transcripts| FTS[(Tantivy FTS)]
    end

    subgraph MCP Search Pipeline
        Query[Agent Query] --> BM25[BM25 Text Search]
        Query --> Dense[Vector Similarity Search]
        BM25 --> RRF[RRF Fusion]
        Dense --> RRF
        RRF --> Rerank[Cross-Encoder Reranker]
        Rerank --> Output[Final Context to Agent]
    end

Pipeline: BM25 probe → Parallel BM25 + Vector → RRF fusion → Reranking (hybrid mode) → Score blending

CLI (development & debugging)

# Index anything
recallforge index ./photos ./docs
recallforge index ~/Movies/demo.mp4
recallforge index ~/Recordings/standup.wav   # requires standup.srt/.vtt/.txt/.transcript.json
recallforge index ~/Documents/roadmap.pptx

# Search any modality
recallforge search "whiteboard diagram from last meeting"
recallforge search --image ./photos/whiteboard.png
recallforge search --video ~/Movies/demo.mp4

# Watch a folder for changes (auto-index)
recallforge watch start ~/Documents --collection docs
recallforge watch list
recallforge watch stop ~/Documents

# Status
recallforge status

RecallForge auto-detects MLX on Apple Silicon, PyTorch elsewhere.

Python API

from recallforge import get_backend, get_storage
from recallforge.search import HybridSearcher

backend = get_backend()
storage = get_storage()
backend.warm_up()

# Index
storage.index_document(
    path="notes.md",
    text="My notes about AI...",
    collection="my_docs",
    model="Qwen3-VL-Embedding-2B",
    embed_func=backend.embed_text,
)

# Search
searcher = HybridSearcher(backend=backend, storage=storage, limit=10)
results = searcher.search("artificial intelligence")
for r in results:
    print(f"[{r.score:.3f}] {r.title}")

Configuration

Variable	Default	Description
RECALLFORGE_BACKEND	auto	auto, mlx, torch
RECALLFORGE_MODE	hybrid	embed, hybrid
RECALLFORGE_MLX_QUANTIZE	4bit	4bit, bf16
RECALLFORGE_STORE_PATH	~/.recallforge	Storage directory

Full references: docs/ENV_VARS.md, docs/MEMORY_POLICY.md, and docs/RUNTIME_BUDGETS.md

For alpha/beta testers:

recallforge flags
recallforge crash-report --output recallforge-crash-report.json --message "what happened"

Testing guidance, feature-flag defaults, and the opt-in crash-report policy live in docs/ALPHA_BETA_TESTING.md.

Project structure

src/recallforge/
├── backends/
│   ├── mlx_backend.py    # MLX 4-bit/bf16 (Apple Silicon)
│   └── torch_backend.py  # PyTorch (CUDA/MPS/CPU)
├── storage/
│   └── lancedb_backend.py # LanceDB + Tantivy FTS
├── cache.py              # LRU query cache with index-version invalidation
├── search.py             # Hybrid search pipeline (BM25 + vector + RRF)
├── server.py             # MCP server (26 tools, stdio + HTTP/SSE)
├── documents.py          # PDF/DOCX/PPTX extraction
├── video.py              # Frame/transcript extraction
├── audio.py              # Transcript-first audio ingest
├── feature_flags.py      # Alpha/beta feature flag registry
├── diagnostics.py        # Local-only crash report helpers
├── watch_folder.py       # Folder monitoring with dedup
└── cli.py                # CLI interface

Development

pytest tests/ -m "not live"    # Unit tests (no model download needed)
pytest tests/ -m live -v       # Integration tests (requires models)

Release Workflow

CI in .github/workflows/ci.yml runs the test matrix, builds distributions, runs twine check, smoke-tests wheel installation, and smoke-tests the HTTP server extra from the built wheel. Tagged pushes matching v* trigger .github/workflows/publish.yml, which publishes to PyPI with trusted publishing.

Before tagging a release, run the repo test suite plus the install/CLI UAT scripts, and if you are on a capable host, run the live integration slice and expanded benchmark. The full checklist lives in docs/RELEASE.md, and routine branch/worktree cleanup lives in docs/GIT_HYGIENE.md.

See CONTRIBUTING.md for full development guidelines.

Attribution

RecallForge is inspired by QMD by Tobi. QMD pioneered the multi-stage retrieval pipeline (embedding, reranking). RecallForge extends this pattern to vision-language with cross-modal retrieval and multi-backend support.

License

MIT License