Memory that just works.

Open-source cognitive memory for AI agents. Local when you want it, shared when you need it.

Website · Quick start · MCP setup · Benchmarks · Discussions

Your agent does not need another place to dump text. It needs memory that can notice when facts changed, explain why something was recalled, keep tenants and workspaces isolated, and remain useful after thousands of writes.

YantrikDB manages that full lifecycle: hybrid recall, temporal decay, consolidation, contradiction tracking, entities and relations, provenance, procedural memory, skills, and explicit maintenance. The core is Rust and the same engine runs embedded, behind MCP, or as a replicated network database.

pip install yantrikdb-mcp

For Codex:

codex mcp add yantrikdb -- yantrikdb-mcp

For Claude Code, Cursor, Windsurf, and other MCP clients:

{
  "mcpServers": {
    "yantrikdb": {
      "command": "yantrikdb-mcp"
    }
  }
}

No cloud account is required. The default store is a local SQLite file and the bundled embedder works without an external model service.

• Memory can be wrong. Contradictory structured claims remain visible and disputed until an operator resolves them.
• Memory changes with time. Recall supports decay, first-mention ordering, revision history, temporal ranges, and point-in-time queries.
• Recall is inspectable. Results include scores and retrieval reasons instead of asking you to trust an opaque top-k list.
• Structure survives sessions. Entities, typed relations, tasks, procedures, triggers, conversations, and skill outcomes live beside semantic memory.
• Isolation is a primitive. Namespaces scope records and cognitive state; the network server adds authenticated tenant databases.
• The deployment path is real. Start with one local file, then move to a server or YRP-replicated cluster without replacing the memory model.

Experimental work lives in yantrik-mind. It explores what a companion can become when typed beliefs, reflection, delegation, and safety all share the same durable substrate.

We publish the awkward results as well as the wins. The benchmark pages include commands, fixtures, caveats, and corrections so claims can be rerun rather than repeated.

• Run the browser Memory Lab
• Read the benchmark ledger
• Whose Memory, Whose Model?
• Skill as Memory, Not Document
• See multi-agent memory handle a stale belief

Use Discussions for architecture and product questions. File engine issues in yantrikdb, MCP issues in yantrikdb-mcp, and server or cluster issues in yantrikdb-server.

The useful feedback is concrete: a memory your agent should have recalled, a stale fact it should have downgraded, a conflict it failed to surface, or a workflow that still takes too much ceremony.