2 of 9

THE PROBLEM

One developer, many agents, no shared memory

OUR REALITY

We run multiple agents: Claude Code for work, OpenClaws for coaching, bots for Garmin health data and workouts, cron agents for HackerNews and YouTube.

ZETTELKASTEN IS NOT ENOUGH

Flat note-based memory leaks context across domains. Your health data bleeds into work context. Meeting notes mix with personal interests. No isolation.

WHAT WE ACTUALLY NEED

An agent-friendly knowledge vault where agents in one domain can learn together, but never mix context with unrelated domains.

AGENT LANDSCAPE

Claude Code

work, code, PRs

Cron Agent

HN, YouTube, docs

Garmin Bot

health, workouts

OpenClaw Coach

training plans

Meeting Agent

transcripts, notes

Flat Zettelkasten / RAG memory

everything mixes together

with NeoCortex

Work Graph

code + articles + PRs

Health Graph

Garmin + training + coach

Meetings Graph

transcripts + decisions

isolated but selectively shared

3 of 9

ACCESS CONTROL

Different agents. Shared knowledge graphs. Permission-controlled.

work-agent

neocortex-platform project

personal-agent

side-project repository

PRIVATE

work_context

project decisions, architecture, team docs

READ + WRITE

PRIVATE

personal_notes

preferences, habits, personal goals

READ + WRITE

hobby_projects

side projects, experiments, learning

READ + WRITE

SHARED

technical_knowledge

frameworks, APIs, patterns, best practices

WORK-AGENT: READ + WRITE

PERSONAL-AGENT: READ

Work agent cannot access personal_notes or hobby_projects.

Both agents access shared knowledge. Permissions control read vs write.

graph_permissions table // per-agent read/write grants // recall fans out only to permitted graphs

4 of 9

ARCHITECTURE

3 MCP tools. Knowledge graph on Postgres.

AI Agents

Claude Code / OpenClaw / VAPI / any MCP client

MCP protocol (streamable HTTP)

remember

recall

discover

Multimodal Ingestion

text / audio / video / meetings -> Gemini multimodal -> extraction pipeline

3-Agent Extraction

Gemini agents: ontology -> entities -> relations -> knowledge graph

PostgreSQL: pgvector (768-dim) + tsvector (BM25) + graph + episodes | Auth0 OAuth | MCP standard

5 of 9

KNOWLEDGE STRUCTURE

Built to solve our own problem: agents that forget everything between sessions

ONTOLOGY LAYER

node_type + edge_type define the schema of knowledge. Agents discover these first before querying.

GRAPH LAYER

Typed nodes with 768-dim embeddings, full-text vectors, importance scores. Weighted edges track reinforcement over time.

EPISODE LAYER

Raw memories: text, transcripts, video descriptions. Append-only with embeddings and access tracking.

MULTIMODAL INPUT

Audio, video, meetings processed by Gemini multimodal into rich descriptions, then extracted into graph.

Each schema is a complete, isolated knowledge graph with its own ontology, nodes, edges, and episodes.

Developer

HN Article

read

Rust Lang

uses

Dist. Systems

about

Meeting

discussed

YT Video

teaches

Indexes

HNSW vector 768-dim

GIN full-text

B-tree traversal

Trigram fuzzy

Per-Schema Isolation

graph_registry tracks schemas

graph_permissions per agent

RLS on shared graphs

Auto-provisioned on first use

We built this because our own agents kept losing context between sessions. This is not a demo project. We are deploying it.

6 of 9

ONTOLOGIES

Agents that understand what they know before they query

WHAT IS AN ONTOLOGY?

A schema of knowledge: typed entities (Person, Tool, Article) and typed relations (USES, AUTHORED, DISCUSSED_IN). It describes the world an agent can reason about.

WHY IT MATTERS FOR GENAI

Without ontology, knowledge dissolves into unstructured embeddings. With it, agents discover what exists before querying. They ask structured questions, get precise answers.

UPPER ONTOLOGY

Groups knowledge into high-level semantic domains, each with its own ontology. Resembles how the human brain organizes knowledge: work memory, spatial memory, procedural memory are separate but connected systems.

SELF-EXTENDING

When new knowledge doesn't fit existing domains, the classifier proposes new ones. Schemas auto-provision. The ontology grows with you.

UPPER ONTOLOGY

user_profile

preferences, goals, habits

Person, Preference, Goal

technical_knowledge

tools, APIs, patterns

Framework, Algorithm, API

work_context

projects, tasks, people

Project, Decision, Deadline

domain_knowledge

facts, concepts, trends

Concept, Article, Event

auto-provisioned

new domain proposed by classifier

schema created on first match

PER-GRAPH ONTOLOGY

technical_knowledge ontology

nodes: Framework, Language, Algorithm, Library, Pattern

edges: USES, IMPLEMENTS, DEPENDS_ON, ALTERNATIVE_TO

proposed by ontology agent, extended per extraction

Like the human brain: separate systems for spatial, procedural, and declarative memory. Connected but not entangled.

7 of 9

HOW MEMORY WORKS

Biology-inspired, not brute-force RAG

EXTRACTION

Ontology Agent

Discovers and proposes node/edge types

Entity Agent

Extracts facts as typed graph nodes

Relation Agent

Builds edges between entities

Upper Ontology

Auto-routes to domain-specific graphs

RECALL

Hybrid Search

Vector cosine + BM25 full-text + graph traversal

ACT-R Activation

Memories strengthen through repeated access

Spreading Activation

Graph neighbors boost recall relevance

Hebbian Learning + Forgetting

Traversed edges grow; unused memories fade

score = w_vec * cosine + w_text * bm25 + w_recency * decay + w_importance * imp + spreading_bonus

8 of 9

ORGANIZATION READY

Auth0 identity. Per-agent isolation. Shared knowledge with zero leaks.

AUTH0 INTEGRATION

Auth0 OAuth / M2M Tokens

Users authenticate via Auth0. Service agents use M2M tokens. Identity auto-provisions into agent_registry on first access.

PRIVATE SCHEMAS

ncx_alice__personal

search_path isolation, no RLS needed

SHARED SCHEMAS

ncx_shared__technical_knowledge

RLS: Auth0 identity -> PG role scoping

ncx_shared__work_context

Fine-grained read/write per agent

AUTH0 AS IDENTITY BACKBONE

OAuth for users, M2M tokens for automated agents and cron jobs. Roles map to graph permissions automatically.

SCHEMA-LEVEL ISOLATION

Each agent gets its own PostgreSQL schemas. No cross-contamination by default. Shared graphs use Row-Level Security.

AUTO-PROVISIONING

First access creates agent identity, personal graph, and permission grants. Zero manual setup per user.

DOMAIN ROUTING

Upper ontology auto-routes knowledge to the right shared graph. Organization knowledge grows organically.

9 of 9

WHAT WE BUILT

RUNNING CODE

5,400+ lines of Python. 70+ tests. Full E2E pipeline. Auth0 integration. Organization-ready from day one.

INNOVATION

Cognitive memory heuristics (ACT-R, Hebbian, spreading activation). 3-agent extraction. Ontology-driven discovery.

REAL-WORLD IMPACT

Cross-domain recall for developers and teams. One memory for work, learning, and meetings. Your agents finally remember.

MULTIMODAL AGENTS

Gemini multimodal for audio, video, meetings. Auth0 for secure identity. MCP standard. OpenClaw/VAPI-ready.

Zbigniew Tomanek / Ignacy Daszkiewicz / Łukasz Łaszczuk

1 of 9

2 of 9

3 of 9

4 of 9

5 of 9

6 of 9

7 of 9

8 of 9

9 of 9