memory OS
Vault scaffold, daily logs, handoffs, qmd retrieval, source verification, and regression fixtures.
PUBLIC REPO. PRIVATE MEMORY.
All-in-one agent OS that does the glue work for you.
FBT is not another agent. It is the installable memory, routing, guardrail, and health-check layer for people switching between Claude, Codex, Claude Code, OpenCode, Hermes, and OpenClaw while using Obsidian, qmd, Ollama, and local models together. It keeps context moving, decides what belongs in cheaper lanes, and stops every harness switch from becoming a restart.
memory scaffold $ fuckbigtech init --with-memory
created: vault/
created: daily/
created: .handoff/
$ fuckbigtech memory-test quick
checks: vault scaffold, handoff folder, demo fixture
result: pass WHY THIS EXISTS
Your agent changed. FBT keeps the work moving.
Claude taps out mid-task, so you jump into Codex.
OpenCode, Hermes, and OpenClaw do not know what your last harness just learned.
Your second brain does not automatically travel with the agent you are using today.
Local and free models are useful, but they need the same context and guardrails.
Every switch burns time unless memory, handoffs, and routing are shared.
INSTALL SURFACE
Hit the limit? Pull the ripcord.
FBT sets up the local memory layer, finds your agent harnesses, wires boring work to cheaper lanes, and runs a degradation test before you trust it. The public repo ships scaffolding and synthetic fixtures; your real second brain stays local.
nowill not read your private vault without permission
nowill not upload API keys or benchmark fixtures
nowill not launch paid agents without asking
nowill not make Obsidian mandatory
$ # Claude tapped out. Your workflow does not have to.
$ # pre-public: print source install guide
$ curl -fsSL https://fuckbigtech.ai/install | bash
$ # then run the printed source install path
$ cd packages/fuckbigtech
$ npm link
$ # prove the memory loop before you trust it
$ fuckbigtech doctor
$ fuckbigtech route "summarize today's handoff"
$ fuckbigtech memory-test quick
first-run wizard LIMIT DETECTED:
Claude: unavailable
Codex: context pressure rising
Local runtime: ready
Memory: stale
Recommended move:
-> preserve handoff
-> route routine work local
-> restart in another harness CONTROL PLANE
It remembers, routes, blocks, and checks the work.
Vault / Handoffs / Daily Logs / qmd
↓
FBT Memory OS + Router
memorycanonical notes, daily logs, handoffs, source reads
premiumstrategy, judgment, high-stakes writing
local/freesummaries, extraction, first drafts, log scans
browser/toolUI checks, screenshots, local app tests
PRIMITIVES
Four things every agent OS needs.
routing
Classify tasks into premium, local/free, browser/tool, or burst lanes before spend happens.
guardrails
Paid-model blocks, side-effect gates, launchd cost allowlists, and public/private boundaries.
benchmarks
Run degradation tests before trusting changes to memory, routing, skills, or qmd.
CATEGORY CHECK
Not another harness. The OS that runs around them.
FBT sits around the tools developers already use. It gives them shared memory, routing policy, side-effect gates, local/free delegation, and health checks so switching harnesses does not reset the operating context.
Claude / Codex
hosted agent harnesses
FBT keeps the memory loop alive when one harness hits limits and you move to another.
OpenCode
coding agent
FBT is the memory and routing layer around agents, including OpenCode.
Hermes / OpenClaw
agent harnesses
FBT gives new harnesses the same handoffs, vault context, and regression gates.
OpenRouter
model router
FBT routes by workflow risk, privacy, side effects, and cost before model choice.
Obsidian / qmd / Ollama
substrate
FBT uses your vault, retrieval index, and local runtime together instead of pretending they are harnesses.
Gemma 4 E2B
optional offline helper
FBT can expose local/browser help for docs questions, but model downloads stay opt-in.
PROOF
The OS is tested before it ships.
The v0 checks prove the install path, private scaffold, safe package contents, and paid-call guard. The larger memory/qmd/routing benchmark suite stays private until synthetic public fixtures are committed.
8/8
CLI smoke
scaffold checks pass
3 files
Package dry run
zero runtime deps
blocked
Paid calls
policy preview only
0600
Private memory
vault files stay local
Public launch should ship synthetic benchmark fixtures and artifacts; private vault runs stay private.
ORGANIC ENGINE
Search pulls intent. Reels create the demand.
FBT is launching like an open-source wedge, not a static brochure. The public content surface answers limit pain, local-first agent OS questions, tool comparisons, and private AI installation searches, then turns the best hooks into reels and social posts.
Claude limits, Codex limits, OpenCode, Ollama, Hermes, OpenClaw, shared memory, local routing.
Definitions, FAQ blocks, comparison tables, source-backed claims, and clean schema for answer engines.
Every strong article becomes short-form hooks for @promptleaks-style distribution and founder/operator posts.
Accounting, law, healthcare, finance, and government-contractor pages route into the AI audit funnel.
BOUNDARY RULE
Public repo. Private second brain.
FuckBigTech ships the portable memory system: adapters, schemas, vault scaffolds, handoff rules, qmd checks, routing policy, cost ledgers, fixture generators, benchmark runners, and synthetic demo vaults. Your real vault, people notes, handoffs, logs, and traces stay local.
shipmemory scaffolds, handoff contracts, qmd checks, synthetic fixtures, benchmark runner, cost schema
keep privatereal vault, daily logs, people notes, raw traces, credentials
OPTIONAL OFFLINE HELP
Docs questions can stay local too.
The v0 site can ship a static FBT help index. The next step is an opt-in Gemma 4 E2B local helper for "how does this work?" questions that runs on the user's device after they choose to download the model.
Static docs search. Works without accounts, keys, or model downloads.
Gemma 4 E2B browser/local helper. User-triggered download, no hosted key custody.
It answers product/setup questions, not private vault questions unless the user explicitly connects local context.
Switch harnesses without starting over.
Install the OS layer, scaffold the second brain, run the health check, then route boring work out of the expensive lane.