dsh-auto-memory

dsh-auto-memory is a plugin for DeepSeek Harness (dsh) that provides persistent memory for agents. It is described as "Claude Code-style persistent memory for DeepSeek Harness — native, zero servers, zero embeddings, zero setup." The plugin stores memories as ordinary Markdown files under $DSH_HOME/memory/, making them hand-editable, grep-able, and git-friendly. It injects a memory index into every system prompt automatically, with zero tokens when the store is empty, and supports typed memories (user / feedback / project / reference) with workspace and user scope layers to avoid cross-project leakage. The system includes crash and concurrency safety via cross-process locks and orphan recovery, and requires no external services, databases, or embeddings.

The plugin offers tools including memory_write, memory_read, memory_list, memory_delete, memory_prune, and memory_delete_all. It follows a write discipline similar to Claude Code: dedupe-and-update over piling up, absolute dates only, [[name]] cross-links, and feedback memories carrying "Why:" / "How to apply:" lines. Auto-consolidation can run on root-session end to extract durable new facts, and forgetting and eviction policies manage memory lifecycle with metadata such as created/updated/reads and staleAfterDays. Recall expansion resolves [[name]] cross-links one level and attaches linked summaries.

The project includes a deterministic evaluation layer in CI that measures injection budget, eviction behavior, and recall quality. It reports that the injected section stays ≤ 4 KB across 20/50/100/200 memories, and that eviction never misfires in a four-class mixed scenario. Known limitations are documented, such as positional budget truncation under half-budget pressure, with pinning improving probe retention to ≥80%. The plugin has 78 tests, 0 runtime dependencies beyond yaml, and is 15 kB installed. It requires @deepseek-ai/dsh >= 0.1.5-rc.2 (Node ^22.19 || >=24).