Neural Bridge
A personal AI substrate: fourteen specialized agents sharing a markdown wiki memory, reachable from a phone via Discord and Telegram, with a filing gate that defends the shared memory against prompt injection and poisoning.
What this actually is, today
V1 ships and runs. Fourteen specialists are defined in the plugin; thirteen identities are registered in the Discord configuration. Luna and Loid have Telegram bridges, and Loid also participates in Council; Loid is not currently registered in the Discord configuration. The specialists read each other's notes, hand off to each other, and emit structured GitHub actions (file an issue, comment, label, close, recruit a new agent). Sessions flush to dated daily logs. A filing gate promotes (or rejects) candidate concepts before they reach the shared wiki. A weekly lint pass re-checks the wiki for drift.
The whole thing runs locally on a Mac Mini under launchd. There is no cloud infrastructure to manage, no service to pay for beyond a Claude Max subscription.
The problem
Most personal AI workflows fragment into point tools — a chat tab, a coding session, a research workflow, each with its own short memory. Every project starts fresh. None of them compound.
Neural Bridge is the substrate where the work compounds.
The substrate, in five layers
1. Agents fourteen .md plugin files, each a specialist
2. Skills inherited from user-level Claude Code settings
3. Transport Discord (mention any agent from any device)
4. Shared state knowledge/ wiki + daily-logs + filing gate + lint
5. Orchestration Discord daemon + senior-pm + cross-agent handoff
The fourteen agents
| Agent | Role |
|---|
luna | Executive assistant: calendar and Gmail via MCP, proactive scheduling, handoffs |
research | Deep reading, citations, threat model write-ups |
teaching-prep | INFO 310 lecture material, slide outlines, exercises |
content | Long-form drafts for the blog and LinkedIn |
social | Short-form posts, X drafts, social copy |
senior-pm | Issue triage, kanban moves, weekly summaries |
recruiter | Designs and provisions new specialist agents |
automation-engineer | Hooks, scripts, daemon work |
security-reviewer | Audits prompts, flows, and PRs for prompt-injection / data-leak risks |
docs-editor | Tightens prose, fixes drift in the wiki |
librarian | Maintains the Luna Master Obsidian vault: INDEX, audits, structure |
echo | Voice-double: keeps Andy's voice profile, flags AI-sounding drafts |
loid | Career strategist on Telegram and Council, backed by the Synapse DB |
ux-designer | Look and feel for neural-bridge-blog and other web surfaces |
@ the registered identities in #neural-bridge on Discord. They read the relevant context, respond, and can hand off to each other (including multi-round /squad-discuss sessions).
Memory pipeline
Daily logs are cheap and per-agent. Concepts are expensive and cross-agent. Promoting a daily log entry into a concept article passes through a filing gate that asks one question: PROMOTE, QUARANTINE, or REJECT?
The gate checks for imperative AI-directed language (textbook prompt injection), untraceable claims, self-promotion, concept-worthiness, slug coherence, and adversarial signal in the source. If a candidate concept fails any of those, it never makes the wiki.
Background and threat model: Memory Poisoning in Personal Agentic AI Substrates.
Discord orchestrator
Thirteen bot identities, one daemon, one asyncio loop. Each agent has its own Discord application and Message Content Intent. A separate Telegram bridge carries Luna DMs and Loid (voice and text); Honcho supplies a shared peer-memory card of Andy across all agents. The daemon:
- Routes
@agent mentions to the right specialist
- Loads the agent's plugin definition into the prompt
- Calls
claude -p with the right tool allowlist (Read / Write / Edit / WebSearch / WebFetch — never Bash)
- Extracts a single fenced
```actions block from the reply, validates it, executes via gh
- Posts the response back as the agent's bot
- Tracks per-channel turn budget so cross-agent chains can't run away
senior-pm also exposes slash commands: /pm-task, /pm-summary, /triage, /squad-discuss, /close.
Private specialist observations
Moonbase telemetry is an optional local JSON
snapshot of the discovered registry and actual model-subprocess execution.
It is disabled unless NB_AGENT_TELEMETRY_PATH names a private file outside
Git checkouts. Registry-only export shows every defined specialist as
unobserved, without starting a bot. Writer freshness, gateway connectivity,
and model activity remain separate; this is not a task-completion feed or the
application-level Fleet heartbeat. No operational endpoint is added.
Repo map
.claude-plugin/marketplace.json plugin marketplace declaration
plugins/neural-bridge-core/ the core plugin
agents/ fourteen specialist .md definitions
hooks/ session_start, user_prompt_submit, session_end, flush, wiki_recall, guards, schema
scripts/ compile (filing gate), lint, discord_bot/
discord_bot/ daemon, mention routing, GitHub actions
launchd/ persistence under launchd
knowledge/ the LLM-maintained wiki
agents/ per-agent session notes
concepts/ cross-agent concept articles (filed via gate)
quarantine/ concepts the gate refused (with reason)
index.md always-loaded starting point
daily-logs/ per-agent session summaries
decisions/ ADRs
docs/ STATUS.md, lint reports, audits
Setup
This repo ships as a Claude Code plugin marketplace plus a Discord daemon.
- Clone the repo and install Claude Code.
- Install the plugin:
/plugin marketplace add andy-herman/neural-bridge
/plugin install neural-bridge-core@neural-bridge
- (Optional) Open the repo as an Obsidian vault for graph view + backlinks.
- (Optional, for the Discord orchestrator) Create one Discord application per agent (thirteen today), enable Message Content Intent on each, store the tokens in macOS keychain (
security add-generic-password ...), populate scripts/discord_bot/agents.json with each application's client ID, then ./scripts/launchd/install.sh to register the daemon with launchd.
Build journal
docs/STATUS.md for chronological progress. Posts about the build and the threat model live at neural-bridge.dev.
License
MIT — see LICENSE.
Attribution
See ATTRIBUTION.md for credits and prior art (Karpathy's pattern, Cole Medin's claude-memory-compiler, AgentPoison and PoisonedRAG threat-model research).
