The
Close your terminal, reboot, come back tomorrow. Normal AI agents forget everything. They burn minutes re-reading old logs, hit rate limits, and freeze up right when you need them.
AegisBrain keeps a permanent notebook. Say “pick up where I left off” and it pulls your exact rules, past fixes, and project decisions in under 25ms. No re-reading. No wasted tokens.
Standard agents crawl 50MB of raw text logs across disk (13,100ms). Aegis recalls exact symbols instantly in 24ms via SQLite WAL.
No more hitting the HTTP 429 Cliff. Injects only 380 tokens of exact symbol context instead of 54,200 raw dump tokens.
Claude: “Context restored from snapshot #042 in 21ms. Continuing authentication middleware: all 286/286 pytest invariants passing. Ready to edit.”
Aegis runs directly in SQLite WAL mode. Your RAM overhead stays negligible, eliminating CPU throttling and host machine freezes.
When terminal sessions terminate, restore points guarantee 100% context survival on your next prompt.
Why pay $20 per million tokens for boilerplate? The expensive model thinks. Free local models on your machine do the typing. You keep the savings.
Hover over any tier layer to inspect the division of labor and real-time laser signal paths.
Offload 92% of verbose boilerplate to free local workers. Reserve precious frontier cognition exclusively for architecture.
Leave an AI agent running overnight and wake up to a frozen terminal. AegisBrain fixes that. A silent Windows Scheduled Task checks every 5 minutes and restarts your agent after sleep, reboot, or crash.
AegisBrainDaemonKeepAlive triggers every 5 min to ensure daemon (:8765) is running.Walk away from your desk and still run the show. Send voice notes, check code diffs, and get pinged the second tests pass, all through Telegram on your phone.
Free local models write fast but they cut corners. Four filters run on every output: strip the filler text, block security holes, sandbox the tests, and loop until zero bugs remain.
Four automated safety filters catch bad code, strip AI apologies, and test every line in an isolated sandbox before touching your git tree.
Every single changed line must directly trace to what you asked. For small edits under 15 lines, changes happen surgically with zero heavy tool overhead. No premature factories, no single-use abstractions, and no speculative flexibility.
// ❌ Rejected (Over-engineered wrapper)
class AbstractConnectionPoolManagerFactorySingleton {
constructor(config: GenericProviderInterface) { ... }
}// ✓ Approved (Karpathy Simplicity)
export function connectDb(url: string) {
return new Database(url);
}See how AegisBrain wires your terminal, memory database, local models, and phone notifications into one system that actually works.
Real architectural pipelines detailing how memory, multi-agent orchestration, and production safety gates work together.
Without a system, coding agents randomly guess, forget decisions, burn through expensive API tokens, and break your repository.
AegisBrain connects your tools into one seamless engine. A senior brain designs the plan, free local models write the heavy code, safety gates test every line, and persistent memory remembers everything tomorrow.
You ask for a feature in your terminal or via Telegram voice note.
$ aegis dev "Add OAuth2 refresh token rotation" [T1-BRAIN] Formulated 4 mathematical invariants & test suite [T2-AGENT] Dispatched boilerplate contract to local worker [T3-LOCAL] Hermes-3 generated 124 lines of candidate code ($0.00) [GATE-01] Karpathy simplicity check: PASSED (surgical diff) [GATE-02] Slop scrubber: Removed AI preambles & apologies [GATE-03] AST security scan: Zero shell injections or leaks [GATE-04] Sandboxed pytest: 286/286 passed in 1.4s [RESULT] Clean commit verified. 0 frontier tokens wasted.
system_topology:
orchestrator:
tier1_brain: "Claude-Opus / Gemini-Pro" (Reasoning only)
tier2_leads: ["backend", "frontend", "security", "qa"]
tier3_workers: ["hermes-cli", "ollama-qwen2.5", "opencode"]
storage:
neural_bus: "brain.db (SQLite WAL mode + FTS5)"
snapshot_limit: 400 # tokens per bootstrap
watchdog:
daemon_port: 8765
keepalive_interval: "5m (Win32 Scheduled Task)"One command. Pick your OS, paste it, done. The background daemon and CLI are ready to go.
Deploy the 24/7 background daemon and global CLI with a single command across macOS, Linux, or Windows.
curl -sSL https://raw.githubusercontent.com/PG300604/aegis-brain/main/install.sh | bashAegisBrain is free and open-source under GPL-3.0. Star the repo, file issues, send PRs.
More stars means more developers find out about it. Takes two seconds.
Get the source code and start hacking immediately.
Request a chassis adapter, report an edge case, or propose an invariant.
Submit PRs for new local worker bridges, IDE shims, or benchmark tests.