Use the docs to get live.
Use the research links when you need claim discipline.
Iranti has two different documentation jobs: help you install and use the system, and help you understand what the current evidence does and does not support. This page keeps those jobs separate so evaluation stays clear.
Quickstart
Install the CLI, bring a database, start an instance, and bind a project.
Claude Code guide
MCP and hook setup for Claude Code, including workspace MCP and response-persistence hooks.
Codex guide
Global Codex MCP registration, workspace files, and the memory protocol for Codex sessions.
Control Plane
Download the local web UI. Browse memory, watch agents connect in real time, and resolve conflicts.
Operator reference
Use the CLI for health checks, lifecycle, project binding, key management, and conflict resolution.
Six ideas that explain how Iranti works and why the architecture choices matter.
Entity + key memory
Iranti is strongest when an agent can address the fact it wants directly. That is why entity + key is the core retrieval model.
The Staff
The Library, Librarian, Attendant, Archivist, and Resolutionist divide memory responsibilities into inspectable components with bounded jobs.
Shared versus personal memory
Personal preferences belong on personal entities such as user/main. Project state belongs on project entities. The system now routes those separately.
Conflict and escalation
Deterministic conflicts resolve automatically when the confidence gap is clear. Ambiguous cases should be understood as conservative and escalation-prone.
Recovery and handoff
Iranti helps most when work crosses agent boundaries, process boundaries, or restart boundaries. Explicit retrieval still beats hoping the model remembers.
Operator visibility
Health, doctor, repair, bindings, version drift, and runtime lifecycle are part of the product story, not afterthoughts.
Full competitive suite results (C1–C4) plus internal benchmarks B1–B14. Includes honest weakness disclosure. Starting point for serious evaluation.
20 facts, 40 questions across Iranti, Shodh, Mem0, and Graphiti. Iranti and Shodh 100%. Mem0 80%. Graphiti 57% — entity extraction loses numeric values.
All 20 facts in one namespace. Iranti: 100% accuracy, 20 tok/query, efficiency 5.0. Shodh: 92% but 66 tok/query. Mem0: 80% at 4.44 efficiency.
Iranti deterministically replaces v1 with v2 (entity+key). Shodh returns both. Mem0 80%. Graphiti 40%.
Fresh subprocess recall. Iranti and Shodh: 100%. Mem0: 75%. Graphiti: 57%. All systems persist — differences reflect retrieval quality.
Exact retrieval vs. long-context reading at 2000 entities, ~107k tokens. Null accuracy gap with large token efficiency differential.
Facts written by one agent retrieved by a completely independent process with a different identity. The core persistence guarantee.
The evidence pages cover benchmark methodology, trial data, and exact claim boundaries. Start with the overview, then go deeper on individual benchmarks.
Ready to install?
Iranti is open source and free to self-host. Run npm install -g iranti to get started.