Foundation by Chroma

Quick Take

Foundation is Chroma’s research preview aimed squarely at what its founder calls “the largest unsolved problem in AI today”: agent memory. Rather than storing a raw log of every agent session, Foundation extracts and structures what was actually learned into a self-improving, wiki-like knowledge base, hyperlinked, tagged, versioned, and built with lineage tracking and access control from the start.

What Foundation Does

Foundation learns from agent sessions and maintains what Chroma describes as a self-improving wiki of what an individual or team’s AI agents have learned. It’s built on ChromaDB and Context-1, Chroma’s own retrieval engine, and Chroma states it’s state-of-the-art on the BEAM memory benchmark, a vendor-reported result worth verifying independently before relying on it for a specific use case.

The core capabilities Chroma highlights:

• Self-improving memory built from agent sessions, rather than memory you have to manually curate
• A wiki-style structure: hyperlinked, tagged, and versioned, so entries relate to each other rather than existing as disconnected facts
• Versioning and lineage: tracking how a piece of knowledge changed over time and where it came from
• Concurrency support: multiple agents or sessions writing to the same memory store
• Access control: managing who and what can read or write specific memory

Why This Connects to Provenance-Aware Memory

Foundation’s emphasis on versioning, lineage, and structure over a flat conversation log is a concrete, shipping example of a broader idea worth understanding on its own: provenance-aware memory, memory that records not just a fact but where it came from, when it changed, and how much to trust it. That distinction matters because unverified agent output, saved without any record of its origin or confidence, tends to get treated as settled fact by whatever reads it next. Foundation’s lineage tracking is one architecture built specifically to prevent that.

It also fits directly into the broader pattern behind Ainanza’s Agent Knowledge Base and Knowledge Flywheel concepts: capturing what worked, from real sessions, and turning it into something future agent runs can actually build on, rather than starting cold every time. Where it differs from a simpler context file or a single project’s persistent agent state is scope: Foundation is built to accumulate structured, cross-session team knowledge over time, not just to carry one task’s current status forward.

Key Facts

• Developer: Chroma
• Status: Research preview (announced by Chroma founder Jeff Huber)
• Built on: ChromaDB (open-source, Apache 2.0) and Context-1, Chroma’s retrieval engine
• Core claims: Self-improving memory, versioning, lineage tracking, concurrency, and access control
• Access: Live now as a research preview at trychroma.com/foundation

Honest Limitations

• Research preview, not general availability. Expect the product to change, and don’t treat it as finished, stable infrastructure yet.
• BEAM benchmark claim is vendor-reported. Chroma’s “state of the art” claim hasn’t been independently verified as of this writing; treat it as a starting point.
• Pricing for Foundation itself isn’t fully public. The underlying Chroma database has clear open-source and usage-based cloud pricing; how Foundation specifically will be priced at general availability isn’t yet documented.
• New enough that integration patterns are still forming. How well it fits into an existing agent stack, versus building your own provenance layer on the open-source database directly, is worth evaluating for your specific workflow.

Alternatives Worth Knowing

• Atlaso, a cross-tool memory layer aimed at individual developers working across Claude Code, Cursor, Codex, and similar tools
• Hugging Face, for general model and dataset hosting, a different layer of the AI stack
• Self-built approaches using ChromaDB directly, for teams that want more control than a managed research preview offers

For the underlying concepts, see Agent Knowledge Base, Provenance-Aware Memory, Knowledge Flywheel, Persistent Agent State, and Portable Context.