Implementation Status

Implementation Status

This page maps Karta's current implementation to major memory-layer capabilities.

Summary

Karta already implements an embedded agent memory substrate:

• durable memory notes
• vector + graph storage
• hybrid retrieval
• LLM-assisted write-time enrichment
• typed/weighted graph links
• background dreaming/consolidation
• contradiction objects
• procedural rules
• access-based forgetting
• MCP tools
• SQLite migrations/health
• synthetic retrieval evals

Karta does not yet implement a full Graphify-style codebase/document KG pipeline:

• no tree-sitter AST graph extraction
• no hypergraph model
• no RDF/OWL/SPARQL/Cypher layer
• no Neo4j/Memgraph backend
• no multimodal ingestion
• no generic graph query MCP tool

Capability Matrix

Capability	Status	Notes
Persistent notes	Done	MemoryNote with id/content/context/keywords/tags/embedding/status/provenance/access stats.
Vector search	Done	Lance-backed vector store.
Graph metadata	Done	SQLite graph store for links, dreams, episodes, profiles, rules, contradictions.
Hybrid retrieval	Done	ACTIVATE pipeline: ANN + facts + profiles + Hebbian + BFS + ACT-R + rerank + RRF.
Session sequence memory	Done	follows links + PAS temporal retrieval.
LLM write enrichment	Done	Context, keywords, tags, atomic facts, foresight signals, link decisions.
Deterministic extraction	Partial	Markdown, JSON, YAML, Cargo.toml extractors. No AST/PDF/OCR/audio.
Dream consolidation	Done	Consolidation, deduction, induction, abduction, contradiction, episode digest, cross-episode digest.
Contradictions	Done	First-class lifecycle: open/resolved/ignored; prefer/merge/supersede/dismiss.
Forgetting	Partial	Access-based note decay, deprecate/archive, protected observed/profile/episode notes. Semantic link decay currently returns unsupported warning (links_decayed = null).
Procedural rules	Done	Safe retrieval/prompt actions only; no arbitrary mutation/code execution.
MCP server	Done	add_note, search, ask, get_note, note_count, dream, get_links. Rust API method behind dream is run_dreaming.
OAuth HTTP server	Done	Axum + OAuth/OIDC + bearer/session auth.
Schema migrations	Done	SQLite schema metadata, applied/pending migrations, health checks.
Eval harness	Done	Synthetic memory eval + benchmark docs.
Codebase KG	Missing	No tree-sitter/function/class/import graph yet.
Graph query language	Missing	No Cypher/SPARQL/Gremlin or generic graph pattern query.
Hypergraphs	Missing	Edges are pairwise typed/weighted links.
Formal ontology	Missing	No RDF/OWL/SKOS/JSON-LD.
Multimodal memory	Missing	No built-in OCR/vision/audio ingestion.

System Shape

flowchart LR
  App[Agent or app] --> MCP[MCP / CLI / Rust API]
  MCP --> Karta[Karta core]
  Karta --> Write[Write engine]
  Karta --> Read[Read engine]
  Karta --> Dream[Dream engine]
  Karta --> Forget[Forget engine]
  Karta --> Rules[Rules engine]
  Write --> LLM[LLM provider]
  Read --> LLM
  Dream --> LLM
  Write --> Vector[(Lance vector store)]
  Read --> Vector
  Write --> Graph[(SQLite graph store)]
  Read --> Graph
  Dream --> Graph
  Forget --> Vector
  Forget --> Graph
  Rules --> Graph

Loading

Implemented Memory Types

Type	Karta representation
Episodic	Notes with session metadata, episodes, episode digests, follows links.
Semantic	Context/keywords/tags/atomic facts + semantic links.
Procedural	ProceduralRule objects with conditions/actions.
Reflective	Dream notes, digests, contradictions, inferred notes.
Temporal	Access history, last accessed time, session sequence links, ACT-R BLL.
Prospective	Foresight signals with TTL/expiry.

Main Data Objects

Object	Purpose
MemoryNote	Primary memory unit.
AtomicFact	Fine-grained extracted fact with parent note.
ForesightSignal	Future-relevant signal with validity window.
DreamRecord / DreamRun	Background inference/consolidation output.
Contradiction	First-class conflict object.
ProceduralRule	Safe symbolic rule that modifies retrieval/answer behavior.
Episode / EpisodeDigest	Session/episode grouping and summary.

Retrieval Detail

flowchart TD
  Query --> Embed[Embed query]
  Embed --> Classify[Classify query mode]
  Classify --> A[Anchor: ANN notes/facts/profiles/foresight]
  A --> C[Co-activation: weighted graph neighbors]
  A --> I[Integration: multi-hop BFS]
  A --> T[Temporal: ACT-R BLL]
  A --> PAS[PAS: session sequence walk]
  C --> Pool[Candidate pool]
  I --> Pool
  T --> Pool
  PAS --> Pool
  Pool --> Rerank[Vector/cross-encoder rerank]
  Rerank --> RRF[RRF aggregate]
  RRF --> Answer[Answer synthesis]
  Answer --> Trace[Trace access + Hebbian writeback]

Loading

Storage Detail

Store	Contents
Lance vector store	Notes, embeddings, vector search metadata.
SQLite graph store	Links, evolution history, dreams, cursor, foresight, profiles, episodes, atomic fact metadata, digests, episode links, rules, contradictions, schema metadata. Uses journal_mode=DELETE, not WAL.

Current Best Use Cases

• personal/agent memory
• research memory
• chat/session recall
• project context memory
• contradiction-aware recall
• long-running assistant memory substrate
• MCP-backed memory tool for coding/research agents

Current Weak Use Cases

• codebase graph analysis like Graphify
• formal ontology/semantic-web storage
• graph database analytics
• multi-modal memory ingestion
• arbitrary graph query service