Repository navigation
Roadmap
Karta roadmap is experiment-driven. Main goal: push long-horizon memory quality beyond 90% on BEAM-style tasks by improving write-time organization, retrieval, and background reasoning.
- Day 2 baseline: 51.3%.
- P0 foundation fixes: 53.0%.
- P1 retrieval fixes: 61.6%.
- Target: 90%+.
Add JSONL per-question logs with full answers, rubric scores, retrieved notes, query mode, contradiction injection, timing, and dream count.
Store real source timestamps and turn indexes so temporal/event-ordering queries can reason over conversation order rather than ingestion time.
When a contradiction dream or contradiction-linked note is retrieved, fetch both source sides and inject them into synthesis context.
Classify queries into modes such as Standard, Recency, Breadth, Computation, Temporal, and Existence. Tune retrieval parameters per mode.
Reduce false abstentions by weakening synthesis-level refusal and relying on retrieval/reranker gates plus natural-language uncertainty.
- Episode narrative ordering.
- Temporal chain links.
- Ghost query paraphrase expansion.
- Hybrid keyword search / SQLite FTS5.
- Contradiction magnet links at write time.
- Meta-instruction pinning.
- Two-pass retrieval on abstention.
Every roadmap item should map to a benchmark failure mode and be validated against BEAM or deterministic synthetic regression tests.
Priority gaps if Karta should fully match the Memory Layer / Graphify direction:
- Graphify-style ingestion: repo/document scanner, deterministic extractors, code entities.
- Generic graph query MCP tools: node/neighbor/path/subgraph queries.
- Graph export/import: GraphML, JSON-LD, DOT.
- Hypergraph support: multi-way relation model.
- Formal ontology adapters: RDF/OWL/SPARQL optional layer.
- Multimodal ingestion: PDF/OCR/audio/vision pipelines.