Celebrum grows a permanent personal memory (a neuron/synapse knowledge graph),
learns a Persona Model from it, guards every autonomous action behind a consent-first
Guardrail Engine, and answers through remembering who you are — not just what you typed.
"Not just code — a memory. For my love, Celebrity."
"What we love most, we fear about them most — and it may be must.""A Valuable gift made for a valuable celebrity — no quantifying."
01 Overview
Celebrum is a local-first artificial brain built entirely on the Python 3.9+ standard library.
No compiled dependencies. It runs on Linux, macOS, Windows, Raspberry Pi, and Android (Termux).
It is not a chatbot. It is a memory system with cognition, governance, and identity at its
center. Every change requires approval. Every change is logged. Every change can be undone.
Everything lives in one SQLite file at ~/.celebrum/. Nothing leaves the device
unless you ship a snapshot.
Pure Python. Zero compiled deps. Install with
pip install ., pipx install ., or uv tool install ..
First brain: celebrum init --demo.
02 How it works
Memory Graph
Neurons and synapses stored in SQLite. Tiered, time-decaying recall: core identity memories (2-year half-life) survive while ephemeral details fade (BRIDGE-styled stable updates).
Persona Model
Learns style, values, heuristics, and tone from the memory graph. Evidential persona (PGMem): signals are believed only when linked to preference/decision memories that ground them.
Truth Engine (Satya)
A veracity layer, a truth index, and a contradiction ratio over the graph. Integration pauses when truth degrades.
Guardrail Engine
Six built-in guards: consent, reversibility, right-to-erasure, human approval for high-tier, append-only audit, data minimization. Nothing autonomous happens without passing a guard.
Simulation
Stance + value-conflict inference before any action, so the brain checks "what would I really think?" against its own persona.
Inference as a service
A standards-based MCP server and a zero-dependency web dashboard.
The flow
Observe→Reflect→Propose→Approve→Remember
03 Built on
The brain is a Python package with a clean modular layout. No compiled dependencies.
store.py
SQLite node/edge store + append-only audit
Storage
memory.py
Neuron/synapse MemoryGraph, decay, recall
Memory
persona.py
PersonaModel + appeal/feedback, identity closure
Persona
tensor.py
Feature-hashed PersonalTensor (no plaintext)
Tensor
truth.py
Satya: veracity + truth index + contradictions
Truth
guardrails.py
GuardrailEngine (DPDP/GDPR gates + audit)
Guardrail
simulate.py
Stance / conflicts / appeal reasoning
Simulation
ingest.py
Duet/HTML/web ingestion (consent + SSRF guard)
Ingest
validate.py
The 6-check neural validation harness
Validation
engine.py
Celebrum facade
Engine
cli.py, gui.py, mcp.py
CLI, dashboard, and MCP server
Interfaces
04 Similar research
Celebrum is built on a lineage of research in memory, persona, and cognitive systems. These are the papers and standards that inform its architecture.
BRIDGE
ICML 2026 poster
Triangular fixed-point refinement for long-horizon persona consistency. Provides a Lyapunov-style uniform drift bound for tiered memory updates.
Celebrum implements: Bridge-drift-bound validation check. Tiered memory decay with 2-year half-life for core identity memories.
ID-RAG: Identity Retrieval-Augmented Generation
arXiv:2509.25299, MIT Media Lab
Grounds an agent's persona in a dynamic, structured identity model: a knowledge graph of core beliefs, traits, and values.
Celebrum implements: Identity-grounded recall — queries are recalled through the statement graph of your identity.
PGMem: Tightly Coupled Persona-Memory Graph
arXiv:2608.01708, 2026
Heterogeneous persona-memory graph connecting event and persona nodes through typed provenance and evidence edges. Keeps each persona signal traceable to supporting events.
Celebrum implements: Evidential persona — persona signals are only believed when linked to preference/decision memories that ground them.
Latent Personal Memory
arXiv:2606.20911
Personal tensor memory — a fixed-shape matrix queried through one additional attention head. Feature-hashed, no plaintext.
Celebrum implements: PersonalTensorMemory (PTM) — a ~2 MB feature-hashed personal tensor for latency-critical similarity and offline drift checks.
DPDP Act (India) & GDPR
Regulatory standards
Data protection frameworks requiring consent, right to erasure, data minimization, and audit trails.
Celebrum implements: Guardrail conformance — six built-in guards (consent, reversibility, right-to-erasure, human approval, append-only audit, data minimization).
llm-guardrails
Open standard
Patterns for guarding autonomous LLM actions behind approval gates.
Celebrum implements: Guardrail Engine with tiered consent (low/medium/high/critical). Nothing autonomous happens without passing a guard.
Celebrum's place: While each of these addresses one or two pieces —
memory, persona, guardrails, or truth — Celebrum brings them together into a single,
consent-based personal system. It is one of only a few such integrated approaches
being developed in India.
05 Validation framework
Measured results:celebrum validate runs the research-backed
six-check neural harness. Every run is audited. All six checks pass.
Check
Source
Result
Bridge-drift-bound
ICML 2026 poster (tiered memory drift bound)
Pass
Identity-recall
ID-RAG (arXiv:2509.25299)
Pass
Evidential validity
PGMem (arXiv:2608.01708)
Pass
Tensor memory
Latent Personal Memory (arXiv:2606.20911)
Pass
Guardrail conformance
DPDP s.6 + llm-guardrails
Pass
Recall latency @10k
SRS 3.3
Pass
Overall
6-check neural harness
6/6 Pass
6/6
Checks passed
100%
Pass rate
2 MB
Tensor size
2 yr
Core memory half-life
10k
Recall latency tested
How to run:celebrum validate — runs the full neural validation
harness from the CLI, or from the dashboard's Validate tab. See docs/VALIDATION.md
for the mapping and docs/SRS.md for the full spec.
06 Project analysis
Validation harness results across six research-backed checks.
Validation checks (all passing)
Bridge-drift-bound
Pass
Identity-recall
Pass
Evidential validity
Pass
Tensor memory
Pass
Guardrail conformance
Pass
Recall latency @10k
Pass
Design principles
Dimension
Approach
Source
Memory persistence
Tiered decay, 2-year half-life for core identity
BRIDGE (ICML 2026)
Identity-grounded recall
Queries recalled through statement graph of your identity
ID-RAG (arXiv:2509.25299)
Evidential persona
Persona signals linked to preference/decision memories
PGMem (arXiv:2608.01708)
Tensor memory
~2 MB feature-hashed, no plaintext, offline drift checks
Latent Personal Memory (arXiv:2606.20911)
Guardrails
Six built-in guards, tiered consent, append-only audit
DPDP s.6 + llm-guardrails
Privacy by design
One SQLite file, SSRF guard, right to erasure
DPDP/GDPR
07 Adaptive cyber-immunity
Celebrum's security layer is derived from biological immune systems. It watches, learns,
remembers threats, and updates its own software — always under consent. The repository
includes an adaptive cyber-immunity Dependabot security framework.
Adaptive cyber-immunity: The repository includes a Dependabot security
framework for adaptive cyber-immunity. Every autonomous action passes through the Guardrail
Engine — consent, reversibility, right-to-erasure, human approval for high-tier,
append-only audit, and data minimization.
08 Research landscape
Known active initiatives across the research areas that inform Celebrum.
Initiatives by category
Commercial AI Memory
4+
4+
US Gov Programs
3+
3+
Organoid Startups
3
3
Academic Grants
3
3
India Public + CSR
2
2
EU Programs
1
1
Personal Projects
1
1
India focus
Project
Research Type
Stage
Engram
Commercial Venture
Active — $98M raised (General Catalyst, Kleiner Perkins, Sequoia)
MemTensor (记忆张量)
Commercial Venture
Active — Pre-A+ round led by 东方富海
红熊AI (Red Bear AI)
Commercial Venture
Active — A+ round, ~$420M post-money valuation
UoH + HAL Digital Twin Lab
CSR Initiative
Active — ₹5 Cr grant from HAL
Allahabad University Neurotech
Government Grant
Active — ₹2.5 Cr from DST
Tata Consulting Engineers CDT
Commercial Platform
Active — built on NVIDIA, pilots with NHSRCL, Torrent Power, PowerGrid
BrainTwin-AI
Academic Research
Published — Brain Sciences MDPI, 2026
BITS Pilani CDT
Unfunded Academic
Published — Taylor & Francis, 2024
PTM (Chennai)
Independent Research
Preprint — Zenodo, 2025
Celebrum
Personal Project
Active — pure Python, zero compiled deps, CC BY 4.0
Observation: Commercial AI memory is the most active category globally.
India's public research presence in cognitive twins remains small relative to global activity.
Personal conviction drives what institutions do not.
09 Principles
◆Local-first. Everything lives in one SQLite file at ~/.celebrum/. Nothing leaves the device unless you ship a snapshot.
◆Consent-based. Every ingest requires explicit consent. Nothing changes without approval.
◆Transparent. Every change is audit-logged (append-only) and reversible. Right to erasure is wired in.
◆Labeled. AI output is always marked as simulated.
◆Yours. Built to remember, not to replace.
10 The code
The web dashboard is served by celebrum gui at http://localhost:8477.
It is zero-dependency and works offline. The repository layout: