Skip to main content
Meerkat Recall is a Rust knowledge system for applications and AI agents. It keeps source material, extracted claims, and a structured graph together, so a reader can ask both “What do we know?” and “What evidence supports it?” Applications connect through HTTP. Agents use MCP tools to ingest evidence, search, navigate relationships, and read cited context. Spaces and attribute-based access control restrict what each caller can see or change. Recall can serve any application or agent stack through those interfaces. In Meerkat, Core runs agents and Meta manages the application and its agent memory lifecycle. Recall adds shared knowledge with sources, relationships and access rules. You can use it independently or add it to an existing Meerkat application. See the Meerkat introduction for component choices. Recall 0.9.1 is available with exact-pinned Meerkat 0.9.0 dependencies. Start with binary installation and service setup; consult the release notes for supported platforms and known limits.

​
Follow one piece of evidence

Suppose a project note says, “Alice owns the Project Cedar test plan.”
  1. A source records where that information came from. A document revision preserves the rendered evidence used for extraction.
  2. The pipeline can identify entities such as Alice and Project Cedar, and extract assertions linked to spans of that evidence.
  3. Validated artifacts populate the graph. Truth maintenance tracks support and competing values for eligible subject-predicate slots.
  4. Retrieval and Living Wiki projections expose this knowledge with citations and freshness information, subject to the reader’s authorization.
If a later note names a different owner, the useful result is more than a new string in a database. Meerkat Recall can retain the competing claims and the evidence behind them. Resolution depends on the predicate, policy, and available evidence; a selected value is not a guarantee that the underlying claim is true. Read the knowledge model for the distinctions between evidence, claims, graph records, and derived views.

​
Start with the experience you need

Try it locally

Run the credential-free demo, then ingest and inspect your own synthetic evidence.

Understand the model

Learn the vocabulary through one example before exploring individual record types.

Build an integration

Authenticate, scope requests, handle asynchronous work, and paginate HTTP results.

Connect an agent

Choose an MCP profile, discover available tools, and follow a complete workflow.

​
What to expect

Extraction is asynchronous and model-driven. Accepting an ingest queues work. It does not mean extraction has completed, or that a model will produce a particular entity or relationship. Inspect processing status and evidence. Evidence and derived knowledge have different lifecycles. Reprocessing can change extracted results or projections while document revisions retain the evidence they were built from. See documents and assertions. Storage is an operational choice. SurrealDB is the default durable backend. The managed embedded elephant-db runtime synchronizes acknowledged commits to a write-ahead log when a persistent destination is configured; explicitly ephemeral library/test stores remain volatile. The deployment guide explains persistence and runtime configuration. This is pre-1.0 software. Use documentation from the same revision as your server. The architecture overview explains component boundaries, including Meerkat’s role in model execution and the MCP transport.