SmritiTantra
A multi-tenant memory layer that grounds AI agents in an org's own verified data and rules instead of model-generated guesses.
Overview
A multi-tenant memory layer that grounds AI agents in an organization's own verified data and rules instead of model-generated guesses, eliminating hallucinated metrics and formulas across agent workflows.
Architecture
A four-plane memory model (data, concept, episodic, and user) built over three shared substrates: bindings, governance, and an event ledger. This separates fast-changing data from curated business meaning, learned experience, and personal context.
Recall Engine
A sub-300ms, LLM-free recall engine fuses graph spread, learned routing, dense retrieval, and lexical matching over a compiled snapshot, returning structurally complete, citation-backed grounding packs for any query.
Governance & Auditability
- Agent-level access control via a 13-verb capability model plus a sensitivity-label lattice, enforced ahead of ranking
- Fully event-sourced and reversible: every write is an append-only ledger event, and any belief can be traced, audited, or unlearned on demand
Learning Loop
A self-improving learning loop with anti-feedback-loop safeguards (propensity logging, Thompson sampling, presentation-bias correction) and proactive gap detection that turns blind spots into evidence-cited questions.
Exposed via HTTP API, MCP server, SDK, CLI, and web console, usable from any agent framework, including Claude Code.