Multi-agent network
Router, Smart orchestrator, and specialists for search, email, tasks, documents, research, images and video, commissioned through one delegation tool.
A multi-agent LLM exocortex.
An AI extension of memory and thinking, living in Slack and Telegram. It remembers, thinks in the background, and answers with context. A solo project in daily use.
Python·GCP·Claude / Gemini / OpenAI / Grok·13,000+ tests + CI
01 / Memory
Every conversation produces facts: preferences, events, decisions, principles. Alek-Core extracts them in the background, deduplicates them against what it already knows, and uses them in every future answer. The more you talk to it, the less you have to explain.
02 / The cycle
A message arrives via Slack or Telegram. A router triages it: complexity, tone, what to search for.
Memory and web search enrich the request. A model tier matching the difficulty answers: cheap for simple, strong for complex.
A background agent turns the conversation into long-term facts. The next conversation already knows them.
03 / Components
Router, Smart orchestrator, and specialists for search, email, tasks, documents, research, images and video, commissioned through one delegation tool.
Multi-vector retrieval with Reciprocal Rank Fusion, versioned facts with full history, number-aware deduplication (64 kg ≠ 66 kg).
Claude, Gemini, OpenAI and Grok are swappable per agent and per user, with complexity-driven model tiers.
A phone call and a web call with a voice companion that shares the character and memory of the text assistant.
Gmail is indexed passively; a daily review triages the inbox into an HTML report.
Memory exposed to Claude as a custom connector (experimental).
An opt-in index of your cloud drive: documents, scans and photos are described once at indexing, so search itself needs no LLM call.
Named procedures the assistant loads on demand. Drafted from chat, inactive until you confirm the save.
04 / Engineering
Hexagonal architecture enforced by 30+ AST-based tests, not by documentation.
13,000+ tests; contract tests validate the payloads sent to each LLM provider SDK instead of mocking the model.
Decisions are recorded: 50+ RFCs and 100+ decision records.
Prompts are data: assembled from versioned tokens at runtime, no inline fallbacks.