Open Beta • personal AI • private memory • trusted knowledge
The model thinks.
KESPA knows.
A personal AI assistant that remembers what matters. Talk with KESPA, work through ideas, build projects, share files, and give it context over time. Useful knowledge can be remembered and brought back later, so every conversation does not have to start from zero.
KESPA is currently in Open Beta.
By using KESPA, you agree to the
Terms of Use
and acknowledge the
Privacy Policy
and
Cookie Policy.
live platform data
Phase:
Open Beta
RTX 3070 8GB
Qwen2.5-Coder 7B
0
Retrieval cards
77
Trusted knowledge records
0
Training events
0
Research releases
78
NexLedger attestations
47
Assistant responses
Live KESPA pulse
Built to improve as its knowledge grows.
These numbers are pulled from KESPA’s live data, so they update as the system changes. They reflect real usage, research, knowledge activity, and ledger records.
Latest activity
auto refresh
Intelligence lifecycle
Latest trusted knowledge
Waiting for live state…
Knowledge promotion activity will appear here.
Latest manual training
Waiting for live state…
Training lifecycle activity will appear here.
Latest NexLedger proof
Waiting for live state…
Canonical proof activity will appear here.
Research pipeline
From gap to reusable intelligence.
Pending knowledge candidates0
System-policy approvals0
Natural research requests pending0
Delivered research requests0
Public research entries13
Public research artifacts262
Trust boundary:
background model output does not become trusted knowledge automatically.
Research must still pass the evidence, verification, policy, promotion, and indexing lifecycle.
KESPA Community
See how the community is using KESPA.
Rankings are based on real KESPA usage, with benchmark and test traffic left out. Only users who choose to make their profile public are shown here.
Loading community…
KESPA Score
What we’re researching
Can continuously maintained, provenance-preserving external intelligence shift the task-performance / compute frontier of resource-constrained models, and do those gains survive when the reasoning model changes?
Research, knowledge, and engineering
Public records are separated by what they actually are.
Research reports experiments and benchmarks. Trusted Knowledge is maintained evidence-backed intelligence. Technical updates document implementation changes. KESPA does not present one category as another.
Latest public research
KESPA-NOTE-030 - Private-Safe Image Editing and Deterministic Transformation Routing
KESPA v7.0 introduced a first-class privacy-aware image_edit path after private routing correctly failed closed when no eligible provider could receive the source image. The implementation authorized the local image bridge, added deterministic transparent-background processing, provider fallback strategies, stronger edit lineage/provenance, and safer error handling while keeping control in Admin/brain_config/provider routing and requiring no Brain runtime modification. Production validation remained pending.
13 research entries
262 artifacts
technical_note
Trusted Knowledge
Key Facts About Life Purpose
What is your purpose?
3 knowledge releases
Evidence-linked
Lifecycle maintained
Frozen V1 benchmark
Verified intelligence materially improved the same small local model on the frozen internal evaluation.
The verified-intelligence condition recorded 54.22 quality, 1.68 s mean latency, and 0.1085 Wh measured GPU energy. The result is preliminary internal evidence, not universal model superiority.
Hit@1 54/54
Gold recall 79.17%
Frozen evidence
Canonical archive health
Loading research archive…
The homepage reads the canonical KespaAI-Research manifest and keeps a last-known-good snapshot. A temporary repository/sync failure no longer silently turns real research counts into zero.
13 entries
262 artifacts
HELP KESPA GROW
Help us keep KESPA moving forward.
KESPA is built independently by NexLabs Studio. Contributions and partnerships help cover the real costs behind the project, including compute, research, infrastructure, and ongoing development.
Support KESPA directly.
Make a one-time contribution of any amount to help fund continued development, research, and infrastructure.
Support KESPAWant to help in a bigger way?
We’re open to hardware, compute credits, infrastructure support, research collaboration, and other partnerships that can help KESPA grow.
Partner with usInterested in KESPA or NexLabs Studio?
Get in touch for more information about the project, our research, the technology behind KESPA, and where we’re taking it next.
Investor informationSystem architecture
KESPA keeps its intelligence outside the model.
The model is only one part of the system. KESPA keeps its long-term knowledge, memory, evidence, permissions, retrieval, and provenance in layers that can continue working even if the model changes.
01 / Knowledge lifecycle
Information has to earn trust.
New information starts as evidence and candidate claims. From there, it goes through verification, policy checks, conflict handling, indexing, and provenance before it can become trusted knowledge.
Evidence
→
Claims
→
Verify
→
Policy
→
Trusted
→
Retrieve
02 / Model
The model focuses on reasoning.
KESPA gives the model the context it needs when it needs it, instead of expecting one model to memorize everything.
03 / Privacy
Your private data stays private.
Personal or private context does not automatically become part of public research, shared knowledge, or the public retrieval system.
04 / Research
Fill real gaps in knowledge.
When KESPA finds something it does not know, the research system can look for public sources, evaluate the evidence, verify new claims, and package the result for review.
05 / Provenance
Every layer has a job.
MySQL stores the authoritative data. Chroma handles retrieval and can be rebuilt when needed. NexLedger records tamper-evident proofs for important system events and research records.
06 / Roadmap
Build something useful first. Expand from there.
NOW
Open Beta
Real users • reliability • privacy • research telemetry
→
NEXT
AR-30 + MM-24
Artifact recall • multimodal evidence • customer validation
→
BUILD
Evidence Store MVP
TXT • PDF • image • spreadsheet
→
GROW
1K trusted claims
Then lock in a longer-term benchmark
Current phase:
Open Beta