First product · Open source
Kalairos
Versioned, time-aware memory engine for long-running AI agents.
Your AI agent forgot what it knew yesterday. Kalairos is a memory engine that remembers changes over time — so agents can ask what is true now and what was true before. Versions every fact automatically. Detects contradictions. Tracks provenance. MIT-licensed.
quick-start.js
const kalairos = require("kalairos")
await kalairos.init({ embedFn: myEmbedder })
await kalairos.remember("User prefers dark mode")
await kalairos.remember("User now prefers light mode")
// Time-travel: what was true a week ago?
const past = await kalairos.queryAt(
"theme",
Date.now() - 604800000
)What it gives agents
Versioned facts instead of overwritten embeddings
asOf queries for historical recall — query any timestamp
Contradiction detection between memory versions
Trust scoring with provenance and source chains
Memory poisoning defense — 5/5 attack classes defended in benchmark
Local-first runtime — JSONL on disk, no cloud database required
Bring-your-own embedder (OpenAI, Cohere, local, or any function)
MCP server included — works with Claude Code, Cursor, and other MCP clients
Why not just use a vector database?
Vector databases store embeddings. Kalairos stores changing memory with version history, provenance, trust, and time-travel. They solve different problems.
CapabilityVector DBKalairos
UpdatesOverwrite or duplicateAutomatic versioning
HistoryNot trackedFull version trail with deltas
Past truthCannot answerQuery any timestamp with asOf
ContradictionsInvisibleDetected across versions
ProvenanceNot trackedSource chain on every fact
Poisoning defenseNot modeledTrust score + contradiction flag + recoverable
DeploymentUsually cloud-firstLocal-first, MIT-licensed
More Krishna Labs products are coming.
Kalairos is the first piece of a broader platform for agent memory, runtime, and applied AI systems.
Talk to us