AI memory for teams — persistent, searchable, auditable knowledge

Correct Once.
Remembered Forever.

Store your company's context in a single, queryable system. Improve AI accuracy, reduce repeated prompting, and keep an auditable history of decisions.

See How It Works
65%
Token Savings
85%+
Retrieval Accuracy
100%
Correction Persistence
<3s
Query Latency (p95)
The Inspiration

Built the Way Your
Brain Actually Works

Neuroscience doesn't store knowledge in flat files. Neither should AI. CortexBrain mirrors the four memory systems that make human cognition extraordinary.

Neural Mapping Active
Diagram showing AI memory model connected to documents and Slack, inspired by human brain neural pathways
Hippocampus Ms Semantic Memory

Your hippocampus converts experiences into long-term knowledge relationships. CortexBrain's Neo4j knowledge graph does the same — entities, relationships, and meaning persist forever.

Prefrontal Cortex Ma Active Memory

Your prefrontal cortex holds active working context and decides what's relevant right now. CortexBrain's Redis spreading activation does the same — relevant knowledge lights up, irrelevant fades away.

Sensory Cortex Mr Raw Memory

Your sensory cortex stores raw perceptions before meaning is extracted. CortexBrain's vector embeddings capture raw document similarity — a fallback when the graph hasn't connected the dots yet.

Metacognition Mmeta Audit Layer

Your brain knows what it knows — and what it doesn't. CortexBrain's metacognition engine tracks confidence, salience, and a full version history of every knowledge correction.

Synaptic

Knowledge stored as relationships,
not flat documents

Activated

Relevant context lights up
on demand, then fades

Correctable

Mistakes are patched in-place
like neuroplasticity

Self-Aware

Knows what it knows —
and what it doesn't

System Diagnostic

Your AI Assistant Has
Amnesia

Enterprise teams using LLMs for knowledge retrieval face three compounding failures that get worse as you scale.

Critical

The Statelessness Tax

Every conversation starts from zero. When your senior engineer corrects the AI — "the port is 3000, not 8080" — that correction vanishes the moment the session ends.

metrics.log
weekly_re_corrections 15-30
Warning

Context Cost Explosion

Traditional RAG stuffs entire documents into LLM prompts. Costs scale linearly. Accuracy drops as context grows — the "Lost in the Middle" phenomenon.

cost_analysis.log
cost_complexity O(n)
Failed

The Accountability Gap

When AI gives a wrong answer causing an outage, there's no audit trail. Nobody can answer: "What data did the AI use? Who corrected it last?"

compliance.log
audit_traceability NULL
SOC 2 HIPAA ISO 27001
Memory Loss Detected
Illustration of AI memory loss — corrections and context disappearing between chat sessions
Field Reports

"We re-correct our AI assistant on the same wrong port number three times a week. It's like training a new employee who forgets everything overnight."

SR

Senior SRE

200-person SaaS company

"Our AI spend tripled in six months. Every query ships 40 pages of context and the model still hallucinates because the answer was buried on page 23."

VP

VP of Engineering

Series B FinTech startup

"After a compliance audit, we were asked to prove which data the AI used for a critical decision. We couldn't. That's when we realized RAG alone isn't enterprise-grade."

CO

Chief Compliance Officer

Regulated healthcare platform

"We built our own memory layer. 6 months and 3 engineers later, it still can't handle corrections without reindexing the entire knowledge base."

TL

Tech Lead, AI Platform

Enterprise logistics company

Competitive Analysis

Everyone Is Trying.
Nobody Has Solved It.

The AI memory space is growing fast, but every existing solution has critical gaps for enterprise use.

Market Matrix
Capability RAG Mem0 Zep Letta Cognee
CortexBrain
Persistent Memory ✗ ✓ ✓ ✓ ✓ ✓
Knowledge Graph ✗ Partial ✓ ✗ ✓ ✓
Corrections Persist ✗ ✗ ✗ ✗ ✗ ✓
Full Audit Trail ✗ ✗ ✗ ✗ ✗ ✓
Confidence Scoring ✗ ✗ ✗ ✗ ✗ ✓
Version History ✗ ✗ ✗ ✗ ✗ ✓
Smart Context Selection ✗ ✗ ✗ Partial Partial ✓
Self-Learning ✗ ✗ ✗ ✗ ✗ ✓
CortexBrain: 8/8 Capabilities
RAG 0/8 Mem0 1/8 Zep 2/8 Letta 2/8 Cognee 3/8
Every existing tool treats memory as a feature. CortexBrain treats it as the product.
Five Architectural Breakthroughs

Not Another RAG.
A Complete Knowledge System.

Every feature is architecturally distinct from traditional RAG. Here's what makes CortexBrain fundamentally different.

01

Corrections Are
Permanent, Not Temporary

When you correct a RAG-based AI, the correction lives only in that session. Next session? Same wrong answer. CortexBrain's four-step mutation pipeline makes every correction permanent across all future sessions.

Locate
Version
Mutate
Meta-Update
02

O(1) Context Cost
— Not O(n)

Our Spreading Activation Engine uses a neuroscience-inspired algorithm to select only the most relevant subgraph — typically under 2,000 tokens — whether your knowledge base has 1K or 1M nodes.

Standard RAG ~3,500 tokens
CortexBrain ~1,200 tokens
65% savings per query
03

Every Answer Has
a Confidence Score

Every response includes a confidence score, source attribution, and full version history. When the AI isn't sure, it tells you — instead of hallucinating.

High
≥ 0.8
Medium
≥ 0.5
Low
< 0.5
Conflict
Flagged

The System Gets Smarter Automatically

When the knowledge base can't answer a question, CortexBrain generates an answer from general knowledge, then auto-ingests it as new knowledge at lower confidence. Weekly consolidation promotes validated knowledge, archives stale facts, and merges duplicates.

Learn
Auto-ingest gaps
Consolidate
Weekly cycle
Promote
0.6 → 0.75
05

Built for Audit,
Not Bolted On

Auditability is architectural. Every mutation creates an audit log entry. Every node has traversable version history. Every answer links to source nodes with confidence scores.

audit-trail
$ curl /api/v1/nodes/{id}/history
[
{ version: 2, value: "3000",
changed_by: "user:priya",
reason: "user_correction" },
{ version: 1, value: "8080",
changed_by: "system:ingestion" }
]
SOC 2 Ready HIPAA Compatible CSV/JSON Export
How It Works

Four Memory Substrates,
Working Together

Inspired by cognitive neuroscience. Built for production engineering.

M_a — Active

Redis 7+

Working memory. Holds activation scores. Relevant nodes get activated; irrelevant ones decay.

M_s — Semantic

Neo4j 5.x

Long-term knowledge graph. Every fact is a node, every relationship an edge. Version history edges.

M_r — Raw

LanceDB

Vector embeddings of all content. Similarity search fallback when graph can't find a match.

M_meta — Meta

PostgreSQL 16

Self-awareness layer. Tracks confidence, salience, access frequency, and audit logs.

QUERY PIPELINE

What Happens When You Ask a Question

1

Entity Extraction

Key concepts identified from your natural language question using NLP-powered entity recognition.

"auth service" "port number" "deployment"
2

Spreading Activation

Neuroscience-inspired BFS traverses the knowledge graph. Only nodes above the activation threshold are selected — bounding context to ~2,000 tokens.

threshold: 30
3

Confidence Gating

Weighted average confidence computed. Response prefixed with appropriate confidence level.

HIGH 0.92
MED 0.61
LOW 0.23
4

LLM Generation + Source Attribution

Context-constrained LLM generates the answer. Every source node cited with confidence. Continuous learning triggers if knowledge gaps detected.

Answer Delivered

With confidence score, source references, version history links, and optional visual output.

Confidence: HIGH (0.92)
Sources: 3 nodes
Tokens: 1,180
See It In Action
Demo CortexBrain: AI That Remembers 2:03
Podcast 15:44

Curing AI Amnesia

How CortexBrain gives AI persistent, self-correcting memory

CAPABILITIES

Everything You Need for Enterprise AI Knowledge

Knowledge Ingestion

Upload PDFs, Markdown, Slack exports, or connect Git repos. Processed through Cognee's ECL pipeline.

.pdf .md .txt git

Versioned Corrections

Correct the AI and it sticks forever. Full version history for every fact.

v1: 8080 v2: 3000

Confidence Scoring

Every answer tagged high/medium/low/conflicted. Tells you when it's unsure.

0.92
0.61
0.23

Smart Context

Spreading activation selects only relevant nodes. O(1) cost at any scale.

~65%
TOKEN SAVINGS

Full Audit Trail

Complete history: who changed what, when, and why.

SOC 2 HIPAA GDPR

Self-Learning

Auto-ingests from LLM fallback answers. Grows smarter through use.

Query Learn Consolidate Improve

Visual Answers

Ask for diagrams, charts, or visualizations. Gemini generates text + images.

text
image

Admin Dashboard

8 dedicated pages for full visibility into your knowledge system.

Health Queries Trends Workers
USE CASES

Who Uses CortexBrain?

DevOps command center with holographic screens showing knowledge graph and AI chat interface

DevOps / SRE Teams

Incident response + runbooks

  • Ingest runbooks, postmortems, and infrastructure docs
  • Correct during incidents — corrections stick for the next on-call
  • Track which knowledge was used during incident response
  • Reduce onboarding time for new engineers

Engineering Leadership

ROI tracking + knowledge gaps

  • Preserve institutional knowledge when engineers leave
  • Measure ROI: token savings, query volume, correction frequency
  • Identify knowledge gaps: which topics have low confidence?
  • Justify AI spend with concrete metrics and dashboards
65%
Token Savings
1.2k
Queries/Week
94%
Accuracy

Regulated Industries

Compliance + audit readiness

  • Full audit trail for every AI-assisted decision
  • Confidence scoring prevents over-reliance on uncertain answers
  • Version history satisfies SOC 2, HIPAA requirements
  • Self-hosted option for data residency control
SOC 2 HIPAA GDPR

Knowledge Management

Single source of truth

  • Consolidate tribal knowledge from Slack, docs, and code
  • Auto-detect and merge duplicate knowledge
  • Track knowledge freshness and staleness
  • Build a single source of truth that improves over time
Slack + Docs + Code Brain
Head to Head

Standard RAG vs.
CortexBrain

Side-by-side: what changes when you replace a stateless retriever with a persistent, self-correcting knowledge brain.

Benchmark
Comparison diagram — standard RAG stuffing documents into a funnel vs CortexBrain selecting precise knowledge nodes
CortexBrain Wins 7/7
Tokens -65% Accuracy +25% Corrections Persistent
Capability
Standard RAG
CortexBrain
Memory Model
Stateless (per-session)
4-substrate persistent
Context Selection
Stuff everything
Spreading activation
Corrections
Lost after session
Permanent + versioned
Confidence
None
4-level gating
Audit Trail
None
Full (who/what/when/why)
Cost Scaling
O(n) linear
O(1) bounded
Self-Improvement
No
Continuous learning
~3,500
Tokens/Query (RAG)
~1,200
Tokens (Cortex)
60-70%
Accuracy (RAG)
85%+
Accuracy (Cortex)
TECH STACK

Built on Proven Open Source

CortexBrain extends Cognee OSS — not a fork, not a rewrite. Proven foundation, enhanced for enterprise.

FOUNDATION
Cognee OSS

Knowledge graph + ECL pipeline. Entity extraction, classification, linking.

LANGUAGE
Python 3.12
API
FastAPI
FRONTEND
Next.js 16
GRAPH
Neo4j 5.x
VECTORS
LanceDB
CACHE
Redis 7+
AUDIT
PostgreSQL 16
LLM
Gemini / Claude / GPT
JOBS
Celery + Redis
Protocol Active

One Brain.
Every AI Tool Connected.

CortexBrain speaks MCP (Model Context Protocol). Plug it into Claude Code, Codex, Cursor, or any MCP client — your entire team shares one persistent brain.

MCP Hub
MCP integration — multiple CLI tools like Claude Code and Codex connected to a shared CortexBrain knowledge brain
All Channels Active
Claude Code ● Codex ● Cursor ● Windsurf ● REST API ●
Connected

Claude Code

Drop CortexBrain into your MCP config. Every Claude Code session gets persistent, auditable knowledge — zero setup.

.mcp.json
{
"cortexbrain": {
"command": "python3",
"args": ["-m", "cortexbrain.mcp"]
}
}
Active

Codex & Any MCP Client

OpenAI Codex, Cursor, Windsurf — they all connect to the same persistent memory via standard MCP protocol.

terminal
# Start the MCP server
$ python3 -m cortexbrain.mcp
▶ MCP server running on stdio
# Or via REST API
$ curl localhost:8000/api/v1/query
6 Loaded

Built-in MCP Tools

query Search with confidence
remember Store persistently
correct Versioned corrections
search_sources Browse datasets
consolidate Memory consolidation
health System health check
API Gateway Online

Integrate With
Any Application

Full REST API with Bearer auth. Add persistent, auditable AI knowledge to your Slack bot, internal tools, mobile app, or any service.

API Gateway
REST API architecture — central server hexagon connected to mobile, web, chat and terminal clients
9 Endpoints Live
Latency <1s p95 Auth Bearer Format JSON
4 Routes

Core Endpoints

POST /api/v1/query

Natural language query with activation-based context selection

request / response
// Request
{ "query": "What port does auth run on?",
"user_id": "priya" }
// Response
{ "answer": "Port 3000...",
"confidence": "high",
"confidence_score": 0.92,
"sources": [...] }
POST /api/v1/correct

Submit a versioned correction — permanently updates the knowledge graph

payload
{ "node_id": "uuid...",
"corrected_value": "Port is 3000",
"reason": "Updated after migration" }
POST /api/v1/ingest
POST /api/v1/ingest/text
5 Routes

Audit & Management

GET /api/v1/nodes/{id}/history

Full audit trail — every version, every change, every user

response
[{ "version": 2, "value": "3000",
"changed_by": "user:priya",
"timestamp": "2026-03-15T14:32Z" },
{ "version": 1, "value": "8080",
"changed_by": "system:ingestion" }]
GET /api/v1/health
GET /api/v1/datasets
POST /api/v1/consolidation/run
GET /api/v1/workers/status
Bearer token auth on all routes
Response <1s p95 Full OpenAPI spec included

Ready to Give Your Organization
a Brain That Learns?

Stop re-correcting your AI. Stop paying for bloated context. Start building institutional knowledge that persists, improves, and can prove where every answer comes from.

Read Technical Docs

CortexBrain: Because your AI should remember what you teach it.