InnerOS Personal Brain

InnerOS role: Product
Lifecycle: Active
Lineage: Living InnerOS product; Battle of the Personal Brains 2026 validated Cognee, Bright Data, AWS Strands, and Docker Sandbox integrations.

Personal Brain is the cognitive and memory layer of InnerOS: durable context, live perception, governed reasoning, and cross-agent memory.

InnerOS principle: hackathons are validation environments. Reusable capabilities are extracted into maintained products and platform layers rather than treated as disconnected one-off projects.

A sovereign cognitive layer for InnerOS that remembers, discovers, reasons, acts, verifies, and learns across agents.

InnerOS Personal Brain is a living InnerOS product, not a hackathon snapshot. It combines shared long-term memory, live external context, local-first inference, agent orchestration, governed execution, and visible evidence.

The architecture was publicly demonstrated at Battle of the Personal Brains, San Francisco, September 21, 2026. The frozen submission from that event is preserved separately in:

Rafa-Innerchispa/inneros-personal-brain-battle-2026

See docs/PRODUCT_BOUNDARY.md for provenance and repository policy.

InnerOS Personal Brain is a live cognitive loop over the existing InnerOS / Ralphi IA ecosystem. Cognee is the portable shared memory brain, Bright Data is live perception, AWS Strands coordinates reasoning and tools, local Qwen/vLLM keeps inference sovereign, and Docker Sandboxes execute bounded actions only after policy checks.

Winning Story

Most assistants start from zero and isolate each agent in its own context. Personal Brain provides a reusable cognitive layer:

Observe -> Remember -> Reason -> Govern/Act -> Verify -> Learn

The system can:

• remember durable personal/project context through Cognee;
• observe the public web through Bright Data SERP REST or MCP search;
• reason through Strands with direct Cognee memory tools;
• synthesize through local Qwen/vLLM;
• govern actions before execution;
• act in a sandbox when allowed;
• learn verified outcomes back into Cognee.

Two-Brain Architecture

INNEROS / RALPHI LOCAL BRAIN                 COGNEE SHARED MEMORY BRAIN
private ops, tools, infra                    portable graph memory
Mongo/Qdrant metadata                        dataset: inneros-personal-brain
MCP/A2A coordination                         Codex / Cursor / Antigravity / Strands
        |                                                  ^
        | curated safe facts, decisions, outcomes          |
        +---------------- CURATED MEMORY BRIDGE -----------+

Cognee is not a decorative memory cache. It is the central shared dataset that multiple agents can recall from and write to. Ralphi MCP remains the local coordination and operations nervous system; losing that route must not erase the Personal Brain's Cognee memory.

Memory Model

Personal Brain intentionally keeps two memory tiers separate:

• Cognee Shared Memory: curated, durable, reusable knowledge that multiple agents can safely recall.
• InnerOS Private Memory: operational/private context retained under owner control in systems such as MongoDB and Qdrant.

Secrets, credentials, customer-private payloads, and private infrastructure details do not belong in shared semantic memory.

Sponsor Stack

Technology	Role	Proof surface
Cognee	Shared graph memory and cross-agent recall	Product memory, official Cognee MCP, Strands memory tools
Bright Data	Live public-web perception	SERP REST via BRIGHTDATA_API_KEY/zone or MCP search_engine
AWS Strands Agents	Reasoning, tool orchestration, audit trail	Direct Cognee tools plus local model route
Local Qwen/vLLM	Sovereign inference	LOCAL_LLM_BASE_URL, local or AMD on-demand route
Docker Sandboxes	Governed execution	Bounded artifact action with evidence
InnerOS / Ralphi IA	Local coordination fabric	Private MCP/A2A route, not the core memory store

Agent Fabric

The product is designed so compatible agents can share the same curated memory fabric. The target proof is simple:

Agent A learns -> Agent B remembers.

Strands, Codex, Cursor, Antigravity and other compatible clients can consume the same shared Cognee dataset through direct tools, MCP, or supported plugins while InnerOS retains its private operational layer.

Demo Evidence

The live UI keeps the judge path simple: ask a question, watch the trace, and read the source evidence. Each answer shows:

• the Strands route decision;
• Cognee memory hits from the shared dataset;
• Bright Data live public-web results, including result titles and URLs;
• local Qwen/vLLM synthesis;
• Docker Sandbox evidence only when Think + Act is requested;
• any fallback or verified replay label when a live route is unavailable.

Verified paths include:

• Cognee remember -> graph -> recall
• Strands -> local vLLM/Qwen
• Bright Data live public-web search
• Docker Sandbox real isolated execution
• Personal Brain API and Live Cognitive Cortex
• Evidence-gated external actions

Backend proof routes remain available at /api/proof/{mode} for smoke tests and technical review:

• remember: Cognee recall with dataset/provenance evidence.
• observe: Bright Data live search, with verified replay labeled only when live search is unavailable.
• govern: consequential action is proposed, policy blocks it, and the proof says NOT_EXECUTED.
• share: Agent A writes a harmless marker and Agent B recalls it from Cognee.

The exact current runtime state belongs in docs/CURRENT_HANDOFF.md.

Battle of the Personal Brains Provenance

The September 21, 2026 hackathon accelerated and validated four integrations:

• Cognee
• Bright Data
• AWS Strands Agents
• Docker Sandboxes

Those integrations remain useful product capabilities, but ongoing development happens here in the living InnerOS repo.

Frozen event snapshot:

Rafa-Innerchispa/inneros-personal-brain-battle-2026

Source snapshot SHA:

ba70fb87913615f613bacb3599c1f3d8738eea5b

Runtime

python -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --host 127.0.0.1 --port 8230

Server-side environment variables:

• COGNEE_SERVICE_URL
• COGNEE_API_KEY
• COGNEE_DATASET=inneros-personal-brain
• BRIGHTDATA_API_KEY
• BRIGHTDATA_SERP_ZONE=inneros
• optional BRIGHTDATA_API_TOKEN or BRIGHTDATA_MCP_URL
• optional LOCAL_LLM_BASE_URL, LOCAL_LLM_MODEL, LOCAL_LLM_API_KEY

Secrets are never embedded in the frontend or committed to Git.

External actions are evidence-gated. The system must not claim an action ran unless its executor returns evidence. Shared memory must remain curated and non-sensitive, while private InnerOS context stays under owner-controlled infrastructure.

Bright Data Choice

For this demo, use SERP API first. It gives judges a fast and easy-to-explain "sees the world" proof for public search. Browser API is only needed for multi-step browser interaction, and Web Unlocker is only needed for a specific hard-to-access page extraction.

Verification Targets

• Product tests: python -m pytest
• Python syntax: python -m compileall app tests
• Frontend syntax: node --check app/static/app.js
• Git whitespace: git diff --check

The demo should never claim a live route is active unless the backend reported it. Replays, missing OAuth, on-demand vLLM routes and blocked actions are intentionally labeled.