Iranti

Shared memory for AI coding tools — Claude Code, Codex CLI, and GitHub Copilot.

Iranti is a self-hosted MCP server that gives your AI tools persistent, identity-based memory. Facts written in one session are retrievable in any other — across tools, projects, and context resets.

Quick Start

# Install globally
npm install -g iranti

# Run the guided setup (configures database, API key, project binding)
iranti setup

# Start the instance
iranti run --instance local

Then wire it into your AI tool:

iranti claude-setup    # Claude Code
iranti codex-setup     # Codex CLI
iranti copilot-setup   # GitHub Copilot

That's it. Your AI tool now has persistent memory across sessions.

Supported Tools

What It Does

Iranti stores facts as entityType/entityId → key → value triples in PostgreSQL. Any agent that knows the entity and key can retrieve the fact exactly — no semantic guessing, no hallucinated state.

Agent A writes:  project/my-app → deployment_status → "deployed to staging"
Agent B reads:   project/my-app → deployment_status → "deployed to staging" ✓

Facts persist across sessions, context resets, and tool switches. When you restart Claude Code tomorrow, it can pick up exactly where you left off.

Key capabilities

• Exact lookup — retrieve by entityType/entityId + key, deterministic and fast
• Hybrid search — lexical + vector similarity when exact keys are unknown
• Cross-tool sharing — Claude Code, Codex, and Copilot share the same memory
• Conflict resolution — concurrent writes from multiple agents are detected and resolved
• Per-fact confidence — every fact carries a confidence score; low-confidence facts age out
• Session recovery — checkpoint/resume for interrupted work
• User operating rules — define trigger-based rules that surface only when relevant
• File-change recall — agents remember which files changed and why

Staff agents

Iranti is built around four internal Staff components that run alongside the host AI tool. Each Staff member has a specific job, and together they turn the memory layer into an active participant in the session — not just a dictionary the agent reads from.

Attendant agency (what the Attendant surfaces on every turn)

The Attendant runs in three phases — pre-response, mid-turn, and post-response — and returns a structured result each time. Beyond raw fact injection, every attend response carries:

• toolCallGuidance — when the host passes a pending tool call (Read, Grep, Glob, Bash, WebSearch, WebFetch), the Attendant derives entity hints from the tool args and emits a shouldSkip verdict when stored facts already cover the target. Hosts can gate tool execution on the verdict instead of string-matching notes.
• drift — detects when the latest message has diverged from the declared task topic. Emits the driving tokens so the host can surface a confirmation prompt. Suppressed when checkpoint.currentStep starts with COMPLETE — so a finished task does not produce spurious drift alarms as the conversation winds down.
• sessionObjective — derived from the task description or checkpoint continuation, threaded through every attend call as a stable anchor.
• autoCheckpointSignal — fires when pressure has built up (drift, turns-without-write, tool-cost threshold) so the host can checkpoint before the next risky step.
• refinementPass — when the first retrieval pass comes back empty, the Attendant runs a bounded widened-hint retry (max 1 extra observe call) and reports the outcome.
• subTurnLoopPlan — on mid-turn attends, when the host passes a partialResponse of the Attendant's own in-progress assistant output, the Attendant re-scores the partial against memory, harvests novel tokens and entity hints from the text, and fires one bounded extra observe call with the widened hints unioned onto the original ones. Net-new facts are deduped against the pre-retry baseline so repeat hits are dropped. Gated by phase, partial length, a once-per-turn budget, and a novelty check on the tokens. This is refinementPass re-applied on response progress rather than empty initial retrieval — the "most agentic" sub-turn loop from the M-series memo.
• attendantToolPlan — up to three planned follow-up tool calls (search_related, observe_entity, query) derived from brief entities, drift tokens, or the session objective. Deterministic and surfaced, never executed.
• councilConsultationPlan — proposes which peer Staff members the Attendant would consult for this turn (e.g. Librarian for source-reliability on a clear topic, Archivist when the injection surface has multiple low-confidence facts). Proposal only.
• usageGuidance — carries the MANDATORY protocol reminder block. Gated on compliance health: when all counters (turnsWithoutWrite, consecutiveUnusedMemoryInjections, etc.) are zero the reminder is suppressed so well-behaved agents do not pay the injection cost every turn.
• writeNudge — reminds the host to write a fact after substantial activity without a durable write.
• toolResultExtraction — on mid-turn/post-response, the Attendant extracts candidate facts from the tool result so the host can autowrite them.
• responseFileCapture — on post-response, the Attendant scans the assistant's reply for file paths, infers the action (edited/created/read) from the ±150-character context window around each match, and auto-writes project/{id}/file/{basename} facts so file-scoped memory is populated without host involvement. Result carries autowriteBatchId, filesDetected, factsWritten, entities, skipped, and durationMs. Only present on post-response attend calls.

Archivist reasoning budget

Each Archivist scan cycle ends with a bounded, deterministic reasoning pass that emits proposals (never mutations) for the Resolutionist to consider:

• compress — clusters of duplicate entries at the same entityType/entityId/key
• flag_drift — clusters with high confidence spread suggesting disagreement
• demote — stale low-confidence single entries
• review_stale — very old single entries regardless of confidence

Proposals fire as reasoning_proposal_emitted staff events and travel on the ArchivistReport so callers can ship them onward.

Council mode

Staff members can propose consultations with each other before finalising a decision. The Librarian can ask the Attendant for relevance when resolving a conflict; the Attendant can ask the Librarian for source-reliability context on a topic; the Resolutionist can ask the Archivist for pending reasoning-proposal context on an escalation. Consultations are proposed, bounded, and fired as council_consultation_proposed staff events — they are not executed automatically today.

MCP Tools

When connected via MCP, Iranti exposes these tools to your AI tool:

Install Strategy

Iranti uses a two-layer model: one machine-level runtime, many project bindings.

1. Install and set up

npm install -g iranti
iranti setup

iranti setup walks you through:

• Instance creation and database onboarding (local Postgres, managed Postgres, or Docker)
• LLM provider API keys (OpenAI, Claude, Gemini, Groq, Mistral, or local Ollama)
• Project binding

Non-interactive automation:

iranti setup --defaults --db-url "postgresql://postgres:yourpassword@localhost:5432/iranti"

2. Start the instance

iranti run --instance local

3. Bind a project

cd /path/to/your/project
iranti project init . --instance local --agent-id my_agent

This writes .env.iranti with IRANTI_URL, IRANTI_API_KEY, and agent identity. Each agent in a multi-agent system gets its own --agent-id.

4. Integrate with your AI tool

iranti claude-setup    # or codex-setup / copilot-setup

API Keys

# Create a scoped key for one user or service
iranti auth create-key --instance local --key-id my_app --owner "My App" \
  --scopes "kb:read,kb:write,memory:read,memory:write"

# List keys
iranti list api-keys --instance local

# Revoke a key
iranti auth revoke-key --instance local --key-id my_app

SDK Usage

Python (PyPI):

from iranti import IrantiClient

client = IrantiClient(base_url="http://localhost:3001", api_key="your_key")

# Write a fact
client.write(
    entity="project/my-app",
    key="status",
    value="in_review",
    summary="App is in review",
    confidence=90,
    source="my_script",
    agent="my_agent",
)

# Read it back
fact = client.query(entity="project/my-app", key="status")

TypeScript (npm):

import { IrantiClient } from "@iranti/sdk";

const client = new IrantiClient({ baseUrl: "http://localhost:3001", apiKey: "your_key" });

await client.write({
    entity: "project/my-app",
    key: "status",
    value: "in_review",
    summary: "App is in review",
    confidence: 90,
    source: "my_script",
    agent: "my_agent",
});

const fact = await client.query("project/my-app", "status");

User Operating Rules

Rules are trigger-based instructions that surface only when the agent is about to do a relevant task (e.g. releasing, pushing to CI). Unlike project policies which are always injected, rules match against the current context using keyword triggers.

# Create a rule via MCP (iranti_write_rule tool) or the API
# Example: remind the agent to use GitHub Releases instead of npm publish
#   triggers: ["publish", "release", "npm"]
#   enforcement: "hard" (required) or "soft" (guidance)

# List all rules
iranti list-rules

# Remove a rule
iranti delete-rule no_npm_publish

Rules are stored as rule/* entities. During iranti_attend, triggers are matched against the current conversation context — single-word triggers match as tokens, multi-word triggers match as phrases.

Diagnostics

iranti doctor              # Validate database, API key, and provider
iranti status              # Show known instances and project bindings
iranti chat                # Interactive chat shell for sanity checking
iranti upgrade --check     # Check for available updates
iranti upgrade --yes       # Apply updates

Operator-facing CLI help now includes short "what it does" and "use this when" guidance for every command — run iranti --help or iranti <command> --help for details.

Configuration

Environment variables (set during iranti setup or manually in .env):

Uninstall

iranti uninstall --dry-run    # Preview what would be removed
iranti uninstall --all --yes  # Remove runtime + project bindings

Guides

• Quickstart
• Claude Code setup
• Codex CLI setup
• GitHub Copilot setup
• Python client
• Security quickstart
• Operator manual
• Conflict resolution
• Cross-tool handoffs
• Vector backends

Links

• Website
• GitHub
• npm
• Python client (PyPI)
• TypeScript SDK

License

AGPL-3.0-or-later