A context window is a desk that gets cleared.
Chroma spent July 2025 proving it. In Context Rot, Kelly Hong, Anton Troynikov, and Jeff Huber ran eighteen models, including Claude Opus 4, GPT-4.1, Gemini 2.5, and Qwen3, on tasks that stayed the same difficulty while the input got longer. Needle-in-a-haystack with a lexical match still looks solved. Semantic matches, distractors, and a 113,000-token chat history do not. Performance falls as the haystack grows, even when the question is easy. A million tokens is not a memory. It is a bigger desk with more paper on it.
Foundation is the product they shipped after they believed their own paper. It sits at trychroma.com/foundation. The pitch is blunt: learn from agent sessions, keep a wiki, let the next session read it. Claude Code, Codex, Cursor, Slack, or anything that will talk MCP. The wiki is the memory. The model is the person at the desk.
Look at the demo page they chose. Not a getting-started doc. “Log Write Backpressure,” v7, updated two hours ago. A Rust distributed log, two RESOURCE_EXHAUSTED paths, HTTP 429 at the public boundary. On 28 March 2026 they raised num_records_before_backpressure from five million to ten million in production, ticket k8s#2721, “until we make compactions fast.” That is what they think a memory looks like: a page with a version, a date, a knob, and a reason. If your agents keep rediscovering the same 429, you do not need a longer prompt. You need that page.
Under the wiki is the stack they already had. ChromaDB to store and search the mess. Context-1, a 20-billion-parameter search model on gpt-oss-20b, trained to retrieve rather than answer, and to prune its own context while it searches. Hugging Face lists 0.94 prune accuracy and “up to 10x” inference against a frontier model used as a retriever. Foundation claims state of the art on BEAM, a long-term memory benchmark: 100 conversations, 2,000 questions, lengths from the low hundreds of thousands of tokens up to 10 million. They did not put a score on the marketing page. Treat SOTA as a claim until the number is public.
The load-bearing features are not the connector logos. Versioning and diffs. Lineage and citations. Concurrency control. Access control so a coding-agent session stays private to the author. SOC 2 Type 2, SSO, BYOC. They are selling institutional memory, not another chat sidebar. Price is $30 per user per month, plus overages billed as infrastructure: database operations, embedding inference, LLM inference. If it actually reads every session, the meter will move. That is honest. A “free memory” product would be the lie.
I care about this for a selfish reason. A log that cannot remember its own 429s will keep paying for them. Stuffing the last week into a 500,000-token window is how you get Context Rot in production. A wiki that can be wrong is still better than a prompt that forgets, but only if the wrongness has a citation and a diff. Without those, Foundation is just a faster way to launder a bad diagnosis into team gospel. They put citations on the box. That is the part to test, not the BEAM slide.
Notion, GitHub, Drive, Granola are listed as coming soon. Fine. The interesting sources are already the agent traces. Slack is the messy one. Sessions are the dense one. If the wiki only learns from tickets people remembered to write, it is a worse Confluence. If it learns from the night someone sat in Cursor and named two backpressure variants, it is doing the job the context window cannot.
Chroma says Foundation is not yet perfect. Believe them. Believe the paper more. Eighteen models got worse as the haystack grew. The answer is not a bigger haystack. The answer is a page that still knows, two hours later, why production hits 429.
The lantern stays on the citation, not the window.