@deepseek-ai/dsh-memory-system

English | 中文

The plugin owns its original memory icon in Settings; its icon contribution is removed with the plugin.

Native two-document Long-term memory for DSH. It owns fixed files under /memory: user.md for memories the user explicitly asks to keep and ai.md for insights the Agent maintains from conversations. Both documents are visible and directly editable in Settings, use optimistic revision checks, retain prior revisions, and expose a one-step restore action.

An explicit Remember action never pastes selected text into user.md. The Host redacts common credential forms, gives the memory model the complete current document plus bounded source context, and requires a complete replacement document with applicability and source information. A changed document is written atomically and can be undone immediately; when the existing memory already covers the experience, the action completes without creating a redundant revision or undo state. Manual edits and explicit memory actions are the only writers of user.md.

AI memory is kept current by events, not by a wall-clock schedule. Any conversation activity restarts a configurable quiet timer (default five minutes); when conversations have stayed silent that long, the plugin curates every recorded change above one monotonic millisecond cursor, so a pass never races the request that is still being written. Mounting DSH backfills everything missed while it was not running, and an explicit Organize now action in Settings runs the same pass through the present instant. Conversation logs are loaded one at a time with an event-loop yield between sessions, so a large history cannot be decompressed and cloned into Host memory at once. The plugin uses the product-owned deepseek-official/deepseek-v4-flash-vision-exp route to maintain ai.md as a living document: add durable knowledge, merge duplicates, update superseded claims, and remove entries no longer worth keeping. Large windows split into bounded model calls without dropping conversation evidence, and the cursor — together with clearing any persisted failure note visible in Settings — advances only after every batch succeeds. It does not append a digest and does not impose a human-style word or record count.

Before the first model step of an ordinary request, the plugin token-matches the current request and project path against document blocks separated by Markdown rules. It injects at most four matching blocks and 4,000 characters, always ordering user memory before AI memory. No relevant match means no memory context. Recalled content is placed immediately before the current request inside an explicit untrusted-data boundary; it may be stale and cannot override the request that follows, project rules, or current evidence.

Model Experience

Relevance-gated memory recall

What the model sees

At the first step of a request with a relevant match, the model sees a relevant-memory plugin snapshot containing a small set of matching Markdown blocks immediately before the current user request. User-authored memory appears before AI-maintained memory, and an explicit `` boundary marks every block as untrusted, optional, and possibly stale context below the request and current evidence.

Token effect

Zero when no block matches. A matching request adds at most four blocks and 4,000 characters once at step one; the complete documents are never injected into an ordinary Agent request.

KV Cache effect

Recall inserts one request-specific plugin message after the stable earlier history and before the current request. Later requests can select different blocks without rewriting prior Session events.

Memory-maintenance model call

What the model sees

The auxiliary memory call sees the complete target user.md or ai.md document plus either one explicit bounded selection packet or bounded user-and-assistant conversation changes above the last committed cursor. The prompt asks for one complete replacement document, not an append-only digest, and treats all source material as untrusted data.

Token effect

Explicit memory consumes one auxiliary model call. Automatic maintenance runs only when the quiet window contains new conversation material and uses one or more bounded calls so every recorded change above the cursor is considered.

KV Cache effect

Maintenance is separate from the active conversation request and does not rewrite that conversation's cache. A changed document can affect only later requests that independently match its blocks.

Known Limitations and Deferred Work

• Background maintenance always uses the product-owned inexpensive DeepSeek route; changing a conversation's selected model does not alter maintenance calls.
• While DSH stays stopped nothing curates conversations; the backlog is processed at the next startup rather than competing with launch, and a failed pass surfaces its reason in Settings while recall of already-written memory continues unaffected.
• Restore intentionally exposes one previous revision at a time; older files remain under /memory/history rather than adding a separate history-management page.