Two recent updates target a practical problem for AI coding agents: retrieving too much irrelevant context. Claude-Mem stages the search for earlier work, while jCodeMunch’s latest release keeps certain generated .NET files out of a code index. They address different parts of the same workflow, and neither release claims to improve coding accuracy.
Memory search, in steps
Claude-Mem v13.35.0, released Oct. 9, adds mem_search, a tool that reveals stored context in bounded stages. In the default guided mode, an agent first gets a short list of saved-context entries, then asks for context around selected entries, and finally fetches a batch of details. It can stop after an earlier step if the titles or surrounding context answer the question. An automatic mode runs the same bounded steps in one call, matching the query’s words against stored text and ranking the matches; the release notes say it makes no language-model calls.
The design gives an agent more control over how much past material it reads at once. The release also adds save_memory for explicit notes, short replies tailored to a query, and an opt-in watcher for importing Markdown folders. The local worker and hosted cmem.ai MCP service use the same search engine and return the same model-facing text, according to the project. The release describes the retrieval mechanics but does not report an evaluation showing better answers or lower total costs.
Using it requires a compatible agent integration and either the local worker or the hosted connection; automatic capture depends on the host. The repository’s code is under Apache 2.0, but that license does not set the terms for the separate hosted service.
Files jCodeMunch skips from a .NET index
jCodeMunch v1.108.338, also released Oct. 9, changes which files its code-search tool indexes in .NET projects. It skips Visual Studio’s .vs/ directory and prunes obj/ and bin/ when a nearby .csproj, .vbproj or .fsproj marks them as build output. It also skips a NuGet packages/ restore tree when the expected package artifact is present. The release notes explain the marker checks: those directory names can hold hand-written source in other languages and project layouts, so a name-only rule could hide real code.
The exclusions run on full-folder indexing, the file watcher and refresh, but not every route. The release notes say index_file ignores skip rules and index_repo applies the .vs/ exclusion but not the marker-based filters; explicit-path resolution can also bypass them. Existing indexes keep old entries until a full index_folder run. These changes target duplicate or vendored code that can crowd source retrieval.
There is also a use restriction developers should see before adopting it: the repository’s dual-use license allows non-commercial use without charge and requires a paid license for commercial use. The README’s token-efficiency figures come from a benchmark on an earlier build, so this update should be judged on the indexing change rather than assumed cost savings.
