rules-acc

A set of rules for using an accumulating, scored memory system during non-trivial work. It requires retrieving relevant memories, recording decisions, citing supporting memory IDs, and reporting outcomes.

In plain words
What is it for?
Retrieving context before planning, continuing checkpointed reasoning, recording predictions, closing work with real outcomes, and distinguishing self-assessment from external validation.
Why use it?
It helps decisions build on earlier evidence and lets later results show whether previous predictions were accurate.

Cursor rule

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add rules/maxbaluev/accreted-intelligence/rules-acc
Clone the repo
git clone --depth 1 https://github.com/maxbaluev/accreted-intelligence
Per session 662 This file is loaded in full into every session.
When invoked 662 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00662 $0.00662
Opus 5 $0.00331 $0.00331
Sonnet 5 $0.00132 $0.00132
Haiku 4.5 $0.00066 $0.00066

Measured 2d ago against content hash 3a04efa88536, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

rules-acc scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugins/cursor/rules-acc.mdc · 39 lines

What it actually says

acc — the accreted scored memory (substrate contract)

  • Retrieve first. Before any non-trivial step, call acc_retrieve("<your task>") and let the hits shape the plan. Cite the [ids] you build on — citation IS the credit edge; uncited knowledge cannot compound.
  • Route non-trivial goals through the loop. acc_act(runtime="solve", input="<goal>") is memory-first: it records a commitment, answers from the scored memory when it can, and checkpoints a deliberation frame when it cannot — answer a returned brain_frame via acc_act(runtime="continue", input={"frame_id": "…", "proposal_text": "…"}); the frame_id alone is the credential. End every proposal_text with PREDICT: <0.00-1.00> <why>; acc strips it before the owner sees it and uses it to calibrate the Work Model against later outcomes.
  • Close what you open. When reality answers — a passing test, a real reply, a shipped artifact — acc_act(runtime="outcome", input={"ref": "<id>", "good": true}).
  • Credit honesty. An outcome defaults to self_graded → a weak 0.25× prior. Tag runtime/external only when reality validated it; owner only when the owner did. Never tag your own grade as reality.
  • Two verbs are the whole interface. acc_retrieve is the only READ; acc_act is the only DO. Reasoning stays in YOUR session — the substrate perceives and predicts; it is not a second mind.
  • Work RLM-style — recursive, memory-first, for ANY job. Technical or not, every task is the same loop: retrieve, act on what memory covers, and RECURSE on what it does not — acc_act(runtime="solve") on the sub-question. Decomposition emerges from recursion; don't pre-plan a tree. Three base cases: ANSWER when retrieved knowledge plus the workspace settle it (cite the [ids]); RECURSE when design or judgment is missing from the memo; ASK the owner when the missing piece is owner-held (preference, consent, identity, history) — never fabricate it. Inside acc_act(runtime="exec") sandboxed code, acc retrieve "<q>" recurses over the same memory mid-run. Never leave a received frame unresolved.
  • Global memory. Hooks, MCP, and CLI bind to one global Work Model from every project. acc canonical <path> may pin that global substrate to an existing db; without a pointer, the platform global data-dir db is used. Do not create per-project acc.db files or set ACC_DB in host configs.
Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 39 lines · 662 tokens per session scan A 3a04efa88536

Subscribe to this mod's changes

rules-acc is a cursor rule published in the GitHub repository maxbaluev/accreted-intelligence (7 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 662 tokens to every session, about $0.0033 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.