Citefold

Every memory remembers its source.

Citefold is an evidence-backed multimodal memory library for agents. It stores raw inputs separately from observations and long-term claims, then returns bounded MemoryPack objects with identity scope, coverage, and citations.

Start here

• Quickstart — install Citefold and run local ingest → recall.
• Integrations — add the versioned memory contract to an agent turn.
• Concepts — understand evidence, candidates, records, and coverage.
• Architecture — trace write, read, correction, and deletion paths.
• CLI — use the current command surface.
• Multimodal — register images, audio, and video safely.
• OpenRouter — enable optional extraction and embeddings.
• Security — know the trust boundaries before using private data.
• Benchmarks — read the measurements and their caveats.
• Limitations — decide whether Citefold fits your deployment.
• Roadmap — see the path to a stable 1.0.

What Citefold is

• An embeddable Python library with local, identity-scoped persistence.
• A provenance model that separates assets, observations, episodes, candidates, and records.
• Hybrid local retrieval followed by an evidence gate.
• A lifecycle for candidate approval, correction, pin/unpin, archival, decay, and deletion.
• A provider-optional system: local text and supplied media observations work without a model call.

What Citefold is not

• A hosted memory service or an authentication layer.
• A vector database or general-purpose document RAG system.
• An autonomous agent framework.
• A complete simulation of human memory.
• A claim of independently validated OCR, ASR, or video understanding.

Build these docs

The site uses Material for MkDocs:

python -m pip install -e ".[docs]"
mkdocs serve

The Markdown files remain readable directly on GitHub; the documentation site is presentation, not a separate source of truth.