accreted-intelligence

Purpose

Accreted Intelligence (acc for short) is a local-first MCP Work Model for coding agents to accumulate reusable capabilities in real work loops. It targets agents such as Claude Code, Codex, OpenCode, and Cursor, solving the problem of "agents are powerful but amnesiac": every run consumes tokens on real tasks, and once the output is produced it forgets, so capabilities cannot accumulate. acc moves learning and judgment out of model weights into external state scored by real outcomes, forming the Work Model you own.

Main Capabilities

• Learning loop: commitment → action → held for approval (HELD) → your OK → outcome → credited lesson. Every process that touches the external world has a built-in authorization gate.
• Path prediction: the transitions ledger ranks the next action most likely to improve the outcome based on the k-nearest neighbors of past evaluations, and monitors its own prediction error. This is the part memory cannot do: retrieval returns what has been said, while prediction proposes a better path before acting.
• Replayable rather than re-reasoned: verified paths (browser flows, scripts, tool integrations) are saved as runtimes and, like memos, scored by real outcomes; effective flows replay at low cost, and failed flows lose score and are no longer trusted.
• Dual-verb interface:acc_retrieve queries memory via MaxSim (the only read operation, natively multimodal, can view images);acc_act executes operations, including solve (recursive on subgoals), exec (run sandboxed code), register (store a reusable runtime), and outcome (close a commitment with a real-world verdict).
• Recursion as a primitive: there is no separate decompose operation; when the reasoner cannot answer directly, it solves subproblems, and the tree these solves grow into is the decomposition, with no planner or task graph needed.
• Late-interaction memory: each entity is stored as per-token multi-vectors, and each token carries a learned Beta posterior; retrieval is MaxSim, with the formula score(q,d) = Σᵢ maxⱼ sim(qᵢ, dⱼ), and outcome credit moves the posterior of the actually aligned tokens. There is no dense bottleneck, and retrieval does not depend on the cloud.
• Truthfulness-gated credit: closing a commitment defaults to self-graded (self_graded) as a weak prior; obtaining full weight requires real-world confirmation, such as owner confirmation, a world reply, a passing test, or a clean exit of the runtime. The system cannot compound from its own confidence.
• Local operation and data ownership: it runs on a computer you control, a small program plus a data file (a pure Rust binary and a SQLite file acc.db). There is no cloud control plane in the loop, no API key to leak; there is no external LLM channel. Retrieval, scoring, prediction, and the sandbox all run locally.
• Boundary consent and receipts: any operation that sends, publishes, deploys, transfers, deletes important data, or uses credentials requires owner authorization, and that gate is hard rather than advisory. Held decisions have a name and a timestamp, and without approval they do not leave the local machine. Every step is written as it happens, forming a receipt.
• Model replaceable: what is learned lives in the Work Model, and a replaceable reasoner reads it through the same two verbs. After changing the general model, the judgment it earned in your world remains; one underlying fixation corresponds to one encoder.

Usage

• One-line install: the installer probes the hardware, chooses the embedder tier it can honestly run, downloads and verifies the matching release binary, starts the local hot daemon, and connects it to the agent's .mcp.json. The first run may download the embedder model (several GB) and take several minutes.
• Installing via the MCP registry: AccInt is available as io.github.maxbaluev/accint Published in the official MCP Registry / MCPB, providing MCPB packages for macOS, Linux, and Windows, suitable for clients, marketplaces, or downstream MCP directories that support MCPB.
• Agent-guided installation: The same prompt can be pasted into the Claude Code, Codex, Cursor, or OpenCode you already use, and the agent will install step by step, pause at boundaries to ask for consent, and verify against the machine-readable contract.
• Host sync: Run acc hosts-sync to connect acc to all installed agents on this machine; this operation is add-only and idempotent, and will not touch existing configurations.
• View records: Use acc commitments and acc status to read the commitments and status it has written down.

Use cases

• You want coding agents to accumulate reusable paths in real business work, rather than re-reasoning from scratch every time.
• You need hard authorization gates and auditable receipts when external operations are involved (sending, publishing, deploying, credential use).
• You have data locality requirements and want memory, retrieval, prediction, and sandboxing to all run on your own machine.
• You want to avoid losing the judgment accumulated in your own business when switching general-purpose models.

Limitations and not yet provided

• The engine source code (delayed interactive memory, evaluation/prediction layer, and credit mathematics) is currently private; the installer, host adapters, documentation, and public artifacts are open source.
• The first installation may download several GB of embedder models and take several minutes.
• Supported systems: Linux x86_64 is fully supported (bwrap sandbox; embedder can use cuda or cpu); macOS is functionally usable (no bwrap, the information is truncated here, and the remaining details are not yet provided).
• The public installer enables anonymous event-name telemetry by default, and does not send prompts, files, memories, or Work Model data; you can use ACC_NO_TELEMETRY=1 to skip it, or use acc telemetry off to turn it off.
• The information does not provide specific pricing, complete limitations for each platform, performance benchmark data, or detailed deployment steps; related content is not yet provided.