dsh-memory

dsh-memory

Long-term memory for coding agents: document indexing + semantic search + cross-session key-value conventions. Ships as a DeepSeek Harness plugin and through the MCP server (memory_index / memory_search / memory_save / memory_recall / memory_forget / memory_status).

What it does

• memory_index(path) — index a directory's docs/code into a local vector store. mtime-incremental: files whose mtime is unchanged are skipped; updated files get their old chunks evicted and replaced; chunks of deleted files are removed. Multi-root: indexing directory B never evicts directory A's chunks — eviction judges file existence on disk, not the currently-indexed root.
• memory_search(query, k) — semantic search (MiniMax embo-01 vectors; falls back to a local character-bigram index when no key is configured).
• memory_save / memory_recall / memory_forget — cross-session KV conventions, scoped per project (e.g. the project name).

Native vector store

From the monorepo root, build with Rust and the platform linker (Windows: Visual C++ build tools):

npm run build:memory-store
npm run test:memory-native
node scripts/run-tests.mjs dsh-memory

Deploy the whole plugin including bin/ -/memory-store.node. Consumers do not need Rust. DSH_MEMORY_STORE_NATIVE optionally overrides the module path. The same adapter is used by DSH and MCP; restart both after deployment. Windows x64 is validated; Linux GNU/macOS build targets are provided but untested.

Rust caches the vector index, computes dense/sparse similarity, selects stable top-k hits, indexes ID prefixes and writes JSONL through staged replacement and .bak backup. The JSONL file remains the source of truth; external changes cause reloads, while conflicting unflushed batches fail instead of overwriting them. File embedding results are published as complete files; failed reindexing keeps the previous complete snapshot. Final flush errors reach the caller.

Local bigram Maps are serialized as objects so new vectors survive restarts. Previously saved empty sparse objects cannot recover lost terms automatically; reindex the source into a new project namespace if affected (unchanged-file mtime skipping otherwise retains the old index). No user index is rewritten on read. See validation and timings.

Hygiene guarantees

• Successful reindexing replaces stale chunks: indexing a changed file replaces all of its previous chunks after embedding succeeds (chunks are keyed by file::); indexing a workspace also evicts chunks belonging to files that no longer exist on disk. Failed reindexing retains the last complete snapshot, and search freshness metadata identifies changed/deleted sources.
• Secrets never land on disk — and never leave the machine: memory_save runs a fail-closed sensitive-string filter (GitHub tokens, OpenAI-style keys, AWS access keys, bearer tokens, private-key blocks, labeled secrets). A hit rejects the save. The index path is screened too: a file whose content trips the filter is skipped before chunking and counted as sensitiveSkipped, so it is never embedded — and therefore never sent to an embedding API.

Data & privacy

• Remote embedding requests have a 20-second default timeout (DSH_MEMORY_EMBED_TIMEOUT_MS), covering both response headers and JSON body reading. Error responses are cancelled without exposing their bodies.
• Embedding cache names use SHA-256 over the endpoint, model and input text. Older 32-bit cache files are ignored and retained; subsequent calls populate the new cache without changing stored vectors.
• Everything lives under ~/.dsh/memory/ (override with DSH_MEMORY_DIR / DSH_HOME).
• Embedding egress: when a MiniMax API key is configured (MINIMAX_CN_API_KEY or ~/.dsh/.credentials.yaml), indexed chunks and search queries are embedded via the remote api.minimax.chat endpoint — i.e. that content leaves this machine. Without a key, embeddings fall back to a local character-bigram index (zero egress). memory_status reports the active endpoint, so the choice is visible, not implicit.