The sovereign
memory layer for AI.
Smritium is a local-first memory substrate that gives every AI model deep, private context — for enterprises and individuals — without surrendering a single byte to the cloud.
Unified Enterprise & Personal Intelligence Infrastructure (UPII).
100%
chunk hashes reproducible on re-ingest
0
cloud dependencies required to answer
3
fused retrieval signals per query
∞
your data, yours, forever
Backed by India's deep-tech ecosystem
What is Smritium
Cloud memory products rent your own mind back to you. Smritium inverts the architecture.
Most memory and “second-brain” systems upload your corpus to a provider, embed it remotely, and meter your recall by the token. Smritium keeps capture, embedding, retrieval and reasoning on your machine — and treats any cloud model as an optional, swappable accelerator, never a dependency. It is the memory substrate that sits beneath every model and agent you use, so your context outlives any single app and stays under your control.
$ smritium ask "what did we agree on latency?"
Answer: A 6-hour delivery SLA, relaxed to 12 hours during the pilot.
Sources: [1] meetings/sync.md (chunk a3f2c9) · [2] terms.md (chunk 77d01b)
✓ answered offline · 0 bytes left the device
Try the open-source core today — the engine is live on PyPI:
$ pip install upiiA local memory substrate for AI models
Everything a memory layer should do. Nothing it should ever see.
Sovereign by construction
Embeddings are computed locally, vectors live in an on-disk store, and metadata stays in local storage. No byte of your corpus is required to leave the device for the system to work — not as a setting, as an architecture.
Verifiable memory
Every answer traces to its source chunks via content-addressed IDs. When confidence is low, it abstains instead of inventing a citation.
The Context Rehydrator
Recall is a sensor-fusion problem, not a cosine distance. Smritium fuses three orthogonal signals into one ranked context window:
Graceful degradation
No GPU, no API key, no internet? Retrieval still answers, and a local model keeps reasoning working. Offline is a first-class mode, never a failure state.
Content-addressed ingestion
Documents are hashed on content, not path. Same inputs always yield the same memory state — idempotent and reproducible.
Consent-gated capture
Ambient capture routes through an approval inbox before anything enters durable memory. Nothing is remembered without consent.
Portable by design
Lift your substrate off one host and re-land it on another without losing continuity. User-owned is enforceable by exit.
Built for developers
A memory layer you can read, run, and verify.
No black-box cloud. Install the engine, point it at a folder, and query it from your terminal, your code, or an MCP-connected agent. Deterministic ingestion means the same inputs always produce the same memory — so citations are stable, testable, and reproducible in CI.
01
Ingest
content-addressed
02
Chunk
deterministic
03
Rehydrate
3 signals fused
04
Answer
cited · offline
# install the local engine (100% on-device)
pip install upii
# build memory from a folder — idempotent, content-addressed
upii ingest ./notes --recursive
# ask: cited, offline, zero bytes leave the device
upii ask "what did we agree on latency?"Unified Enterprise & Personal Intelligence Infrastructure
Personal context is a primitive, not a feature.
Today's AI stack has models, agents and connectors — but no memory-and-identity substrate that outlives any single app and that a person or an organisation actually controls. UPII is that missing layer: a single, persistent, owner-controlled context graph that sits beneath every model, agent and application.
Whoever holds the richest, freshest representation of a person — or an enterprise — produces the best output. UPII makes the natural owner of that representation the person and the organisation, not the platform. Fire any model without losing yourself.
How Smritium compares
Most “memory for AI” is a cloud you rent. Smritium is a substrate you own.
The mainstream approach ships memory as a metered, cloud-hosted API. Smritium ships it as a local substrate — sovereign, verifiable, and offline by default. Here's how the architectures diverge.
Smritium✓ supported · — partial · ✕ not supported. Reflects typical cloud memory APIs and local RAG toolkits.
The sovereign scoreboard
Others count the memories in their cloud. We count the ones we'll never see.
0+
local memory engines installed — and counting
Every one keeps its memories on the user's own machine.
0
of those memories stored on our servers
Not a policy — an architecture. What never leaves cannot leak.
Who it's for
One substrate. Two kinds of sovereignty.
The Smritium Membrane
The compliance layer your regulators have been asking for.
When something truly must cross to a cloud model, the Membrane detects and redacts PII, shows exactly what is about to leave, requires human approval, and writes a tamper-evident egress audit log. Legal, finance, healthcare and government teams get provable data sovereignty — not a policy PDF.
PII detection & redaction
See exactly what would leave before it does.
Tamper-evident egress audit
Every byte that ever left, provable.
On-prem / air-gapped
Deploy org-wide, entirely inside your walls.
Compliance posture
DPDP & GDPR by architecture, not paperwork.
Measured, not asserted
Verifiability is our benchmark.
We publish the harness, not just the headline. Re-ingest the same corpus and reproduce 100% of chunk hashes — a citation you got last month still points at exactly the same evidence today.
0.958
Recall@10 on a committed labelled set
100%
chunk-hash reproducibility on re-ingest
0
required cloud calls to answer
3
retrieval signals fused per query
Verify, don't trust — the evaluation harness ships with the open-source core.
Why we're building this
Memory is infrastructure. It should belong to you — not to someone else's cloud.
India's AI sovereignty
The intelligence layer of the next decade shouldn't depend on foreign clouds holding the nation's context. Smritium keeps the memory substrate on your own hardware, on Indian soil — a building block for sovereign AI.
Enterprise AI sovereignty
Your institutional knowledge is your moat. It should compound inside your walls — not inside a vendor's model, meter, or terms of service you don't control.
Data privacy by construction
Privacy that is architectural, not a checkbox. Capture, embedding, retrieval and reasoning stay on-device. What never leaves cannot leak.
Government of Karnataka
Backed by the ELEVATE Nxt programme · 2026
A deep-tech grant-in-aid initiative supporting frontier AI research built in India.
Built as research, shipped as product — developed by DataFrontier Innovations with a provisional patent on multi-signal context rehydration, published benchmarks and a reproducibility audit.
smṛti · स्मृति — “that which is remembered.” Memory has been treated as sacred infrastructure in Karnataka for ~3,000 years longer than the cloud.
The road ahead
What's upcoming on smritium.com
The open-source core is live today. Here's what's shipping next as Smritium grows from a local memory engine into a full sovereign substrate for people and enterprises.
Local memory engine — on PyPI
Content-addressed ingestion, the Context Rehydrator (semantic + temporal), 12 CLI commands and a local knowledge graph. Install it today: pip install upii (v0.7.0).
Relational retrieval + ambient inbox
Knowledge-graph re-ranking across people, projects and organisations, plus consent-gated ambient capture that flows through an approval inbox before entering durable memory.
MCP bridge for agents
Give Claude, Cursor and other agents smritium_search / smritium_ask / smritium_list_sources — returning only user-approved, cited chunks. The corpus never leaves the device.
Cmd+K everywhere + encrypted sync
A lightweight overlay puts your memory one keystroke away in any app, with multi-device sync over LAN / encrypted P2P — no cloud in the loop.
Mail & calendar connectors
Consent-gated capture from mail (.mbox / .eml) and calendar (.ics), with debouncing, dedup and delete-handling.
Smritium Membrane — Enterprise
Org-wide egress audit + PII redaction, on-prem / air-gapped deployment, and DPDP / GDPR compliance reporting your security team can actually read.
Answer verification & smritium forget
Abstention when confidence is low, and user-facing verified deletion — remove a memory and prove it's gone, end to end.
Want early access to each drop as it lands?
Early access
Join the Smritium waitlist.
Be among the first to run the sovereign memory layer. Tell us whether you're exploring for yourself or your organisation — enterprise design partners get priority onboarding.
Be the first to reserve your place