01 Cross-model grounding 02 Memory that carries forward 03 Open Factlet protocol

Models know the world. Nora knows you.

Each AI session sharpens the next — measured, cited, audit-ready.

Every session produces patterns, decisions, and fixes — organizational learnings that should carry forward but vanish when the tab closes. Nora captures them automatically and grounds every AI reply in your team's facts — so each answer cites what your team actually knows.

Kernora provides the decision layer for AI-assisted work — grounded in your private data, built on the open Factlet protocol, audited like code, run on your hardware. BYOK · zero telemetry · on-prem for regulated work.

BYOK · Local-first or BYO cloud · Supervised + autonomous · Elastic-2.0
claude code · payments-api · main
Engineering · DEMO
LYou · 11:42

How should I handle Stripe webhook retries here?

NClaude · grounded by Nora

For this codebase, use an idempotency key on every handler f024 and exponential backoff with jitter, capped at five retries f117.

Don't retry without dedup — that caused incident f214 last quarter (double-charge on /refund).

3 facts injected 1 incident referenced 247ms
claude desktop · close-q2-2026 · main
Finance · DEMO
RYou · 09:18

Can we book the Acme renewal in Q2 or does it slip to Q3?

NGPT-5 · grounded by Nora

Slips to Q3. Your ASC 606 policy recognizes revenue only when the contract is signed and the performance obligation has started f312. The Acme MSA was signed 06-28 but service start is 07-01 f298.

Beta Corp had the same shape last year — booked early, restated in audit f178.

3 facts injected 1 precedent referenced 286ms
nora desktop · plant-12 · controlled
Manufacturing · regulated · DEMO
MYou · 04:12

Pump-03 is reading 4.5 mm/s vibration. Within spec?

NGemini · grounded by Nora

At spec — but flag for review. Pump-03's commissioned alarm threshold is 4.5 mm/s, last verified 2026-04-15 f234. ISO 10816 Class II baseline is 4.5 mm/s for this frame size f246.

The same bearing housing failed at 4.7 mm/s in 2024 after sustained operation — incident f091. Schedule inspection within 72h.

3 facts injected 1 incident referenced 198ms

DEMO · sample data · not real customer factlets

−86%
Completion tokens vs. ungrounded baseline (2026-05, N=6 models)
grown_factbook eval, 2026-05
20×
Fewer high-risk shipping recommendations
factlet-ai/evals, N=6 models
≤250ms
Context injection latency, p95
nora_context.py · CSafeLoader
0*
Bytes sent to Kernora servers
*Free/Lite tier · Pro adds BYO S3 sync · verified by tcpdump on install
What we solve

Memory and grounded context for AI-assisted work.

Grounding and memory are the same artifact — a verified body of facts your team owns (conventions, specs, history, decisions) that the model can cite. Nora builds it from your repo and sessions, maintains it as work evolves, and serves it to every agent — solo to company-wide, supervised or autonomous.

01 Grounding · guardrails  — engineering example
01

Stop the model from making things up about your codebase.

Generic AI gives generic advice. Your team has hard-won decisions, regrets, and conventions that aren't in any training set — and they conflict with the obvious answer often enough that "obvious" is dangerous.

Factlet protocol Verified facts Citations · f### Per-fact egress consent

Kernora builds a factbook from your repo, docs, ADRs and tickets — then grounds every reply in your team's facts, forcing each answer to cite what it used. Open standard, audit-it-yourself.

02 Memory · carries forward  — engineering example
02

Stop teaching your AI the same thing every session.

A session ends, the context window dies, the lesson dies with it. Tomorrow morning you'll explain again why the migration script can't run inside a transaction. The third time you explain something is the third time you've wasted that prompt.

Session analyzer Trust ladder Git-tracked lineage Contradiction detection

Every session ends with new facts proposed, verified, and merged into the same factbook. Tomorrow's prompt starts where yesterday's ended.

Factbook · the artifact

A queryable, citable knowledge base — AI-ready and version-controlled with your repo.

A factbook is plain markdown plus a structured index in .nora/ next to your code. The model can't quote a fact without naming it.

How Nora maintains it — a continuous loop, not a one-time import.

01 · Collect

From repo & sessions

Reads code, ADRs, tickets, docs. Watches every Claude / Cursor / Kiro session as it happens — supervised or autonomous. Each captured fact gets an ID, source, decided-on date, and privacy tier.

02 · Analyze

When sessions end

Nora extracts decision candidates, dedups against existing facts, flags contradictions, scores AI Leverage for the session.

03 · Curate

Trust ladder

Facts enter as proposed, auto-promote to candidate after N uses without contradiction, you approve canonical from the dashboard.

04 · Propagate

To every agent

Injected via MCP or system prompt — Claude, GPT, Gemini, or local. Same factbook, every model, every session.

New sessions feed back to 01 · your factbook grows with every session
One click from the decision

Click any f### chip in a reply to open its decision trail.

See what it cites, why your team decided it, what it superseded, who applied it last. Git-tracked alongside your code.

See it in action

Built on the open Factlet protocol — your factbook is portable, your tooling isn't locked to us.

Architecture · deployment topology

One factbook in the middle. Five sources, five surfaces, five guarantees.

Nora maintains the five elements of the Factlet protocol in the middle box — facts flow in from your tools, out to every model you use, with provenance and consent attached.

Factlet protocol v0.9 · W3C PROV-O lineage Live deployment topology
SOURCES FACTBOOK · MAINTAINED BY NORA EMIT FORMATS Repo · code, ADRs, tickets Sessions · Claude / Cursor / Kiro Docs · PDFs · notebooks Slack · Linear · GitHub Auto-memory · Claude / Cursor factbook ON-DEVICE · PORTABLE · PROV-O f:fact statement + ID f:lineage source · supersedes f:consent per-fact egress tier f:confidence trust ladder · applied N f:emit serializer contracts n Nora resident analyst v3.1 schema Claude · system prompt OpenAI · tool result Gemini · context cache MCP server · resource SDK · raw JSON-LD SOURCES → FACTBOOK → EMIT · one artifact, five surfaces
Validated openly

Grounding works. Independently measured.

In an open, pre-registered eval across Claude Sonnet 4.6, GPT-4.1, and Gemini 2.0 Flash (2026-05, N=6 models), supplying the model with a team-specific factbook reduced harmful shipping recommendations 4× and high-risk recommendations 20× — same direction across all three vendors. Nora is the reference implementation that gets your factbook generated, kept current, and into the model's context.

Reduction in high-risk shipping recommendations with a factbook in context.

N=6 models, held-out prompts, blind-judged. Methodology and raw data on GitHub — factlet-ai/evals.

See the methodology + raw data on GitHub →
Claude Sonnet
22×
GPT-4.1
18×
Gemini 2.0
20×
Coverage · the honest gap

Know what your factbook doesn't cover.

A representative-query batch runs nightly. You see which questions your team will ask that your factbook can't ground — and a concrete list of facts to add. Not a score: a punch list.

Coverage today
62%

of 184 representative queries hit FactSignal ≥3 (2026-05).
Up from 42% two weeks ago.

14-day trend
Apr 26 May 02 May 09 · today

Bucketed histogram · 184 queries (2026-05)

28
0 bars · dead
42
1 bar · weak
39
2 bars · thin
47
3 bars · ok
22
4 bars · strong
6
5 bars · grounded

The 0–1 bar buckets are the productive output: 70 queries your team will ask but your factbook can't ground. Click a bucket to see the queries; click a query to see the proposed-fact diff.

Dead-zone drilldown · 0 bars

28 queries · click to propose facts.

Sample · top 5 by ask-count
"How do we rotate the staging API key?"
7 asks · 14d
propose fact →
"What's our retry policy for the billing webhook?"
5 asks · 9d
propose fact →
"Who owns the data-residency runbook for EU customers?"
4 asks · 12d
propose fact →
"What's the SLO for nora_search p99 latency?"
3 asks · 5d
propose fact →
"Do we ship to the React 19 server-component path?"
3 asks · 11d
propose fact →

Verified nightly. Falsifiable. Sharable with your team in one screenshot.

Not another markdown vault

"Why not just use Memory Bank?"

Memory Bank documents what your project is. Kernora captures what your team decided — and why — automatically.

Memory Bank / Roo Code

You write, you maintain, it drifts.

  • — You write and update markdown files by hand
  • — Stores project docs — stack, conventions
  • — Loads everything at session start
  • — Drifts silently when you forget to update
  • — Flat files — human-readable, not queryable
Kernora / Nora

Automatic capture, typed, decayed.

  • ✓ Hooks capture intelligence during sessions
  • ✓ Stores decisions, reasoning, alternatives
  • ✓ Injects only what's relevant to this session
  • ✓ Auto-decay plus nora revise for corrections
  • ✓ Structured graph — queryable via MCP

Memory Bank is a journal you write every day. Kernora is a colleague who was in every meeting.

Get started

One command. Solo to team to production.

Install in 30 seconds for individuals. Sync your own S3 bucket for teams. Deploy as a sidecar for autonomous agents in production. Same factbook, end to end.

CLI · curl

Claude Code · Claude Desktop

Registers Kernora as an MCP server. Restart your agent and the factbook tools appear.

~/code · bash
curl -fsSL https://kernora.ai/install | bash
→ tcpdump audit · 0 packets to kernora.ai
✓ Restart Claude · type /mcp to verify
SDK · Docker · production

Autonomous agents in production

Sidecar to your agent runtime — CI bots, scheduled agents, customer-facing agents. Headless MCP, BYOK secret management, audit log to your S3, decisions auto-promoted by trust ladder. Per-fact egress consent, encrypted at rest, your bucket and keys throughout.

production · docker
# sidecar your agent runtime, no Kernora cloud
docker run -d kernora/agent:latest \
   --factbook s3://your-bucket/factbook \
   --emit-mcp tcp://0.0.0.0:2742
→ Audit log · your S3 · your KMS keys
✓ Ready · agent runtime can call factbook.query
The origin

Why we built this.

I was shipping a product with Claude Code, Kiro, and Cursor — sometimes three agents in parallel. Hundreds of sessions. Thousands of decisions. And the same thing kept happening: I'd solve a problem on Tuesday, and on Friday the AI would suggest the exact approach I'd already rejected. The models got smarter every quarter — but everything I taught them evaporated every session.

The better a session went, the more I lost when it ended. Reusable patterns, architectural rules, hard-won fixes — none of it carried forward unless I manually wrote it down. The METR study showed experienced developers are measurably slower with AI — not because the tools are bad, but because the context cost of re-explaining your codebase eats the productivity gain. Anthropic's own data: engineers fully delegate only 0–20% of tasks despite using AI in 60% of their work. What slows real teams down is missing context, not missing model capability.

The more you work, the smarter your AI gets. Automatically.

Founder

Mihir Choudhary

Bay Area

LinkedIn →

Built products and led teams at Amazon. Built startups. Now building the local memory layer that grounds AI for AI-assisted work — visibility into how developers, managers, and organizations adopt, measure, and benefit from AI tooling.

Get in touch

Open source, commercial, or a thoughtful note — we'd love to hear.

GitHub Issues → hello@kernora.ai