Skip to content
Christian Pojoni edited this page Apr 27, 2026 · 4 revisions

Karta Wiki

Karta is an experimental agentic memory system that thinks, not just stores. It combines an embedded vector store, a graph/state store, LLM-assisted write-time organization, cross-encoder reranking, background "dreaming", contradiction tracking, forgetting, and procedural memory.

Status: experimental research project. APIs, schemas, benchmark numbers, and defaults can change.

Start here

  • Quick Start — run Karta as a Rust library or MCP server.
  • Architecture — write/read/dream paths and storage layers.
  • Configuration — env vars, model setup, storage, reranker, ACTIVATE.
  • MCP Server — OAuth-protected MCP endpoint and tool list.
  • Memory Concepts — notes, links, dreams, contradictions, rules, forgetting.
  • Benchmarks — BEAM 100K status, reproducibility notes, eval commands.
  • Development — local checks, PR expectations, contribution rules.
  • Roadmap — retrieval-quality experiment backlog.
  • FAQ — common questions and failure modes.

Repository

Core promise

Most memory systems store text and retrieve nearest neighbors. Karta tries to organize memory at write time, keep relational structure, reason in the background, detect contradictions, decay stale content, and explain retrieval evidence.

Main capabilities

  • Self-organizing note graph.
  • Dream engine: deduction, induction, abduction, consolidation, contradiction detection.
  • First-class contradictions with lifecycle: open, resolved, ignored.
  • Cognitive retrieval pipeline: ANN, keyword, Hebbian, ACT-R activation, integration BFS, rerank, PAS, facts, foresight, profile channels.
  • Procedural memory: safe rule DSL for query/session/contradiction conditions and prompt/retrieval actions.
  • Forgetting engine: preview/run lifecycle with protected notes.
  • Evidence packet shape for why-retrieved traces.
  • Deterministic extractors: Markdown, JSON, YAML, Cargo.toml.
  • Embedded defaults: LanceDB + SQLite.

Completion Pages

Clone this wiki locally