chroma

Chroma — open-source vector database company; cited in 2026-04-27-llm-training-principles-paths-practices for Context-1, their training work on a self-editing search agent.

What this wiki currently knows

Products / projects referenced

• Context-1 — Chroma’s research project on training a retrieval-augmented agent that learns to prune its own context. The published artifact is the trychroma.com/research/context-1 post.

Techniques attributed

• prune_chunks as a learned policy — context pruning isn’t a pre-baked retrieval heuristic; it’s a tool the agent learns to invoke, with rewards on retrieval-side relevance. Distinct from the more common pattern of fixed top-k retrieval + post-hoc filtering. (meta-harness, eval-grader-reward)
• Mid-trajectory document scoring — the reward signal credits the agent for finding relevant documents along the way, not only for the final answer. One of the article’s three case studies for “reward design fragments in Agent training”.

Stance

Chroma sits next to moonshot-ai (Kimi PARL) and cursor (Composer 2) in the article as the third public reference for non-trivial Agent-training reward decomposition — but they’re solving a narrower problem (retrieval policy) than the other two. The interesting framing is that their solution (prune_chunks as a learned tool) treats context engineering itself as a learnable surface, which is parallel in spirit to Meta-Harness‘s outer loop. Tracked here primarily as a reference point; will grow if more retrieval-as-learnable-policy work lands.

The vector-search substrate Chroma productizes — embeddings + approximate-nearest-neighbor MIPS — is exactly what Weng’s 2023 survey (2026-06-04-llm-powered-autonomous-agents) catalogued as agent long-term memory. Context-1’s prune_chunks is one answer to that survey’s unspoken open question: the 2023 framing optimized how fast to retrieve (LSH/HNSW/ScaNN); the harder problem turned out to be what to retrieve and what to discard.