Privacy and local-memory policy

August 15, 2026 · View on GitHub

English · 简体中文

DSH Passive Memory is experimental software that can retain sensitive personal conversation history. Operators are responsible for deciding which users, agents, and workspaces may share a memory scope.

What is retained

By default, the local SQLite evidence store may retain:

  • direct user text;
  • displayed assistant text;
  • ordinary tool results as evidence-only records;
  • source-linked L1 summaries when episode compression is enabled;
  • content-free recall and memory-tool audit metadata.

Model reasoning, tool-call blocks, plugin-generated recall messages, and output from memory_expand or memory_search_raw are excluded from L0 ingestion. Native DeepSeek Harness session logs remain separate authoritative sources.

Where data is stored

The database path is chosen through PASSIVE_MEMORY_DB_PATH or plugin config. SQLite may create -wal and -shm companion files beside the main database. The repository ignores these files, but filesystem access, backup software, cloud synchronization, and malware scanners remain outside the plugin's control.

When data leaves the machine

  • Lexical storage and retrieval are local.
  • Episode compression sends the bounded episode evidence to the model route already selected in DeepSeek Harness, or to an explicit override.
  • Semantic retrieval is opt-in. When enabled, bounded evidence text is sent for document embeddings and the accepted user prompt is sent for query embeddings.

Provider terms and retention policies therefore apply to those requests. API keys are read from process-environment variables and must not be placed in deployment YAML or committed files.

Isolation and deletion

Retrieval requires an exact user, agent, and workspace scope match. Scope ids are isolation controls, not display labels; reuse shares memory, while changing one creates a separate retrieval boundary.

Release 0.1.0-alpha.1 has no end-user export, selective-forgetting, or retention-period UI. To erase one store, stop DeepSeek Harness, back up anything that must be retained, then delete the configured database and its matching -wal and -shm companions. This does not delete the separate native Harness session logs or copies held by an external model or embedding provider.

Publication fixtures

Real validation conversations, persona documents, generated source maps, and development databases are excluded from the public repository. Public tests use synthetic fixtures. See LOCAL_DATA_POLICY.md for contributor rules.