December 2025 Summaries

Supermemory has introduced a Startup Program for early-stage founders building AI products, aiming to reduce the cost and engineering complexity of implementing reliable memory, context, recall, and persistence systems. Participants receive three free months of the company’s Scale plan, including roughly $600 per month in usage, along with dedicated support, rapid onboarding, and time to build and launch products. The company says the initiative reflects its own experience with infrastructure challenges and support from its community, and is intended to help startups test ideas faster without maintaining custom memory infrastructure. Early-stage companies and prospective founders are invited to apply for the program.

Supermemory introduces code-chunk, an AST-based library for preparing source code for retrieval-augmented generation systems, arguing that character-based text splitting often breaks functions and loses meaningful programming context. Built on the cAST research approach and tree-sitter parsers, the library identifies semantic entities such as functions, classes, methods, imports, signatures, documentation, and parent-child scope relationships, then recursively creates size-bounded chunks at syntactic boundaries using non-whitespace character counts. It enriches chunks with contextual metadata including file paths, scopes, defined signatures, dependencies, and nearby code, intended to improve embedding quality and retrieval relevance. Beyond the research prototype, code-chunk provides overlap options, streaming and batch processing, Effect integration, and WASM support for edge environments, while supporting TypeScript, JavaScript, Python, Rust, Go, and Java. The authors also revised their evaluation setup with same-repository distractor files and intersection-over-union relevance thresholds, reporting stronger Recall@5 and IoU@5 results than fixed-size and alternative code chunkers, as well as reduced time, tokens, cost, and tool calls in a SWE-bench Lite agent experiment using semantic search.