xater · open-source flagship

Хәтер

/xæ'ter/ — Tatar word for «memory». Latin spelling: xater.

A memory layer for digital assistants that belongs to the human, not to the model. A graph with provenance, different decay for different types, emotional weight, decode-time confabulation blocking.

Why

I run ~20 projects in parallel with different teams and domains — from oilfield services to pharma networks and NPOs. Cloud assistants don't remember who Marat Ramilevich is, what OKPD is, or how Alfa-SK differs from Mech-Stroy. After every «new chat» I retell the same things again and again. Хәтер solves this — memory survives sessions and belongs to me.

Architecture: 4 waves

Memory is not uniform. Behavioral data, personal preferences, secrets and episodic events require different rules of storage, indexing and decay.

1
Public projects

Behavior, reference data, business logic. Extracted to the graph normally, indexed in vectors.

2
Self / Feedback / Reference

Personal preferences, feedback, references. NOT ingested into the graph — only into vectors. Reason: feedback is noisy and must not generate entities.

3
Secrets

Passwords, tokens, IDs. An isolated project, marked Sensitive, never passed to graph_extract.

4
Episodic / temporal

Events with a timestamp and emotional weight. Decay by type: hot 7 days, warm 60 days, cold — archived.

Claimed novelty

I worked through prior art — Generative Agents (UIST 2023), MemoryBank (AAAI 2024), HippoRAG (NeurIPS 2024), Google and Apple patents. The concept is 60-70% taken. The clean novelty is a narrow combination:

  1. Provenance-typed graph with different rules per source type
  2. Confabulation blocking at decode-time (not as a post-filter)
  3. Per-type configurable decay with emotional weight

Strategy — not a patent, but a defensive publication: arXiv in 6 months + AGPL repository + a series of posts describing the mechanism. Creates public prior art and reputation instead of $30k at the USPTO.

Roadmap

  • M1 — repo skeleton, first article on Habr
  • M2 — bridge + poller release, English version on dev.to
  • M3 — post «Provenance in LLM memory», paper kickoff
  • M4 — post «FZ-152 and sovereign assistants»
  • M5 — paper draft, external review
  • M6 — arXiv + announcement

Current status

alpha · code locally on VDSina · repository github.com/Rustam260287 (private until M1) · license AGPL-3.0