DSH | Code Memory | Local source context to reduce repeated code discovery

[kobubu]

Unofficial community project, independently developed and maintained. Not affiliated with or endorsed by DeepSeek.

Project: kobubu/Rag-Coding

I built Code Memory (CM), a local JavaScript/TypeScript code-memory plugin for DeepSeek Harness. It gives the agent relevant source and dependencies before the first model step, aiming to reduce repeated glob, grep, and read calls.

CM uses exact file/symbol lookup when the task names its target, and local MiniLM embeddings, LanceDB, and a compiler-derived dependency graph when discovery is needed. It checks source freshness and avoids repeating unchanged whole files already supplied to Harness. Indexes and embeddings stay on your machine; selected source still goes to your provider if you use a cloud model.

Published pilot: three 36-session series on one repository, with six task types and matched no-CM / older-CM / newer-CM arms. The newer CM used 45–66% fewer total model tokens, with 18–37% lower Harness session time than no CM across those series. Independent checks passed 36/36 newer-CM attempts and 35/36 no-CM attempts, pooled across the three series.

These are small single-project pilot results, not guaranteed savings. Setup/indexing time is separate, and all failures are retained. Method, raw results, and limitations. Earlier approaches that did not work.

Integration: native Harness bundle with automatic context, cm_retrieve_context, and a generated project profile. The Windows release pins Harness 0.2.0-rc.2. MCP and optional manual hooks are also available for Codex and Claude Code; the Harness percentages above do not describe those clients.

CM Atlas screenshot: this public repository's src directory, 39 modules and 8,321 structural relationships. The current map UI is in Russian. This screenshot shows code navigation, not a token-saving benchmark.

Try it: download CM-0.9.0-Windows.zip and SHA256SUMS.txt from the release, verify and extract the ZIP, then run:

powershell -NoProfile -ExecutionPolicy Bypass -File .\install_cm.ps1 -ProjectRoot 'C:\path\to\your-project' -Clients harness

English installation guide. Dependencies and model weights download on first install; configure your normal model provider in Harness. No model credits are included.

I would welcome independent tests on other repositories, including regressions: same starting commit, task, model, and checks, with CM on/off. Please include total tokens, elapsed time, failed attempts, and whether the patch actually passes independent checks.