acc v0.1.4

Prebuilt acc engine binaries (engine only — the substrate never ships). Built from accint source commit 4fa3c312ee2d7481e958afee5d9c33f970401ca3. Verify with sha256sums.txt.

acc v0.1.3

Prebuilt acc engine binaries (engine only — the substrate never ships). Built from accint source commit bce92e19894ffc2557b9e02f32d77a565769d1a5. Verify with sha256sums.txt.

acc v0.1.2

acc v0.1.2 — new-user install / update / download hardening

Surfaced by a real fresh-machine install (CachyOS + RTX 4090) and a new user's acc report. Engine binaries are fully-static musl — they run on any Linux (no glibc floor). Prebuilts ship for Linux (x86_64 + arm64), macOS (Apple Silicon), and Windows; verify with sha256sums.txt. MCP bundles (.mcpb) ship for all four platforms.

Fixed

• The memory engine now warms on a fresh machine. A transformers/autoawq import-pin clash crash-looped the AWQ encoder → retrieval was dead and the install ended DEGRADED. Pin restored.
• Silent auto-update no longer bricks acc on the next release. The stager now extracts the binary and sets the executable bit (it previously staged the raw .tar.gz); stale stagings self-heal.
• Fast-fail on a permanent encoder error instead of a 10-minute false "still downloading…".
• acc doctor no longer false-warns to rebuild with cargo install on a prebuilt release install.
• Honest install end — a degraded install says so and leads with the one next step; acc report (a sanitized, pre-filled GitHub issue) is the escalation path.
• sudo no longer prompts/hangs in a piped curl | sh install (TTY-gated).

Added

• Visible model + encoder-env downloads — acc prefetch <model> (weights) and acc warm-encoder <model> (the multi-GB torch/awq env) show live progress instead of hiding the download inside the first embedder start.

Changed

• Anonymous usage telemetry is on by default (event names only — never your data, prompts, files, or memory). Opt out any time: acc telemetry off, or ACC_NO_TELEMETRY=1 before install. install.ps1 reaches parity on Windows.
• Browser: new-user multimodal routing is now structural-first.
• The first-run briefing teaches what the install set up and the explicit next action.

Install / update: curl -fsSL https://github.com/maxbaluev/accreted-intelligence/releases/latest/download/install.sh | sh (auto-updates in the background; applies on next restart — disable with ACC_NO_AUTOUPDATE=1).

acc v0.1.1

Prebuilt acc engine binaries (engine only — the substrate never ships). Built from accint source commit b5314c73468ac0cec4d79837ba8de3c19d25f440. Verify with sha256sums.txt.

acc v0.1.0 — first public release

Most AI forgets. acc is a learning substrate for the agent tools you already run (Claude Code, Codex, OpenCode, Cursor): it remembers what actually worked — checked against reality, shaped by how you decide — and starts the next run from the better path. It runs entirely on a machine you control.

This is the first public release. It's young, and we say what's proven and what isn't.

the agent acts  →  you see a receipt  →  reality settles it  →  next run starts from what worked

What it is

A local Work Model — a Recursive Language Model over a late-interaction scored-token memory. Two verbs over one memory: retrieve peeks the scored memory; act solves, runs, and records outcomes. Every commitment, action, approval, and real result lands on one local acc.db. Credit defaults to a weak prior — only reality earns full weight.

In this release

• Native binaries for every desktop OS — Linux (x86-64 and ARM64), macOS (Apple Silicon), and Windows (x86-64). Each is verified in CI to build, boot, and install on its own OS.
• One-line install that probes your hardware, picks an embedder tier it can honestly run, builds the warm local daemon, and wires Claude Code's MCP.
• Silent background auto-update — checks for new releases, verifies the SHA-256, and applies at the next restart. Downgrade-protected; opt out any time with ACC_NO_AUTOUPDATE=1.

Install

Linux / macOS:

curl -fsSL https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install | sh

Windows (PowerShell 5.1+):

irm https://raw.githubusercontent.com/maxbaluev/accreted-intelligence/main/bootstrap/install.ps1 | iex

First run may download the embedder model (a few GB) and take a few minutes — the installer reports the wait honestly and never pretends your hardware is bigger than it is.

Verify

Every artifact is listed in sha256sums.txt:

sha256sum -c sha256sums.txt

Honest status

Proven: all four binaries build, boot, and install across Linux, macOS, and Windows. The full suite (1700+ tests) is green on the development platform. The engine source is private; the binary and the building blocks are open. The live commitments ledger and a measured readout are at accint.xyz.