The core profile is about 2,964 tokens of schemas; the all profile reaches 87 tools and roughly 16,212. Rather than making you choose once, discovery is a tool: search the catalogue for a name, then load that one schema. The deletion tools show similar care — batch_forget refuses to run without a filter, and the descriptions point you at lowering importance instead of deleting when what you want is suppression rather than removal. Note that this package is only the client; the memory server is a separate thing you have to be running.
The MCP front end for a Dakera memory server you run yourself. Fourteen tools load by default; the rest of the catalogue stays behind a discovery pair so a session does not begin by spending its context on schemas.
- Memories stored with an importance weight and tags, where 0.8 to 1.0 marks the ones that must survive decay (dakera_store).
- Four retrieval shapes for four different questions: top-k semantic recall, semantic search constrained by tag or memory type, BM25 keyword search for exact terms like error codes and IDs, and a single-pass hybrid where vector_weight slides between the two (dakera_recall, dakera_search, dakera_fulltext_search, dakera_hybrid_search).
- Filter-based listing by tags, importance range, time window, type or session, for when semantic matching is not what you need — at least one filter is required (dakera_batch_recall).
- Results expanded along knowledge-graph edges one to three hops out, and a graph built from a seed memory by embedding similarity to see how a concept connects to what is stored (dakera_recall with include_associated, dakera_knowledge_graph).
- Sessions that group everything stored during one run under a shared context, opened with metadata such as task type and closed with a retrievable summary (dakera_session_start, dakera_session_end).
- Deletion in both shapes, each with a guard: by explicit ID or tag, and by filter with at least one filter required so a full wipe cannot happen by accident (dakera_forget, dakera_batch_forget).
- Structured extraction of entities, topics, key phrases and a summary from free text, through a provider chain that falls back to a local GLiNER model when no external provider is configured (dakera_extract).
- The rest of the catalogue on demand: search it by keyword or tier and get names with one-line summaries, then load the full input schema only for the tools you are about to call (dakera_discover_tools, dakera_load_tools).
A Dakera memory server of your own — the MCP process is a client, not the store. Run it from the published Docker image, or use the deploy repository's compose file for persistent storage, then point DAKERA_API_URL and DAKERA_API_KEY at it. The client itself installs however you prefer: npx or a global npm install on Node.js 18+, Homebrew, cargo, a Docker image, or a pre-built binary for macOS, Linux or Windows. DAKERA_MCP_PROFILE decides how many tools appear in tools/list, and it can also be set per request.
One command plus a key — npx @dakera-ai/dakera-mcp, then supply credentials