working-memory

A task memory file that stores decisions, open questions, changed files, and success criteria outside the agent's chat context.

In plain words
What is it for?
Use it to record the goal, assumptions, file changes, chosen approach, and ongoing decisions for one coding task.
Why use it?
It helps the agent keep consistent decisions during long tasks or after the conversation context becomes crowded.

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/ag-dmitry/composer-enhanced/working-memory
Clone the repo
git clone --depth 1 https://github.com/AG-Dmitry/composer-enhanced
Per session 26 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 662 The whole file, excluding the scripts and references it only reads on demand.
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 $0.00026 $0.00662
Opus 5 $0.00013 $0.00331
Sonnet 5 $0.00005 $0.00132
Haiku 4.5 $0.00003 $0.00066

Measured 2d ago against content hash 33db1a6ee310, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

working-memory 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.

agent-protocols/working-memory.mdc · 71 lines

How it starts

The opening of the file, as written. The whole thing — 71 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Working Memory

Why

A cheap model has a finite context window. When it fills up, early decisions (the "why" of a design choice) quietly evaporate. Later in the same task, the model re-examines those choices from scratch — and may reverse them inconsistently, creating the flip-flop / oscillation failure mode.

The working-memory file acts as durable, structured external state. It survives context compaction and model resets. Writing to it forces the model to distil its decisions explicitly, which itself improves quality (externalizing = understanding).

File

One file per task at .cursor/agent-memory/<task-slug>.md.

  • Template: templates/task-memory.md.
  • Not committed to git — add .cursor/agent-memory/ to .gitignore.
  • Lifecycle: create in phase 1, update throughout, discard or archive when done.

Sections

Section When to fill
Goal Phase 1
Success criteria Phase 1
Non-goals Phase 1
Complexity Phase 1
Open questions Phase 1 + updated continuously
File ledger Phase 2 (path → summary + relevant line ranges)
Decision log Phase 3 (chosen approach + rejected alternatives)
Iteration log Phase 4-7 (what changed + gate outcomes)
Red-team findings Phase 5
Self-report Phase 6 on pass

Rules

  • Distil, do not dump. Write the key decision, not the full reasoning chain.
  • Always update before a context boundary (compaction, end of chat, model reset).
  • Before the iteration-3 re-plan checkpoint, re-read the entire file to restore full context.
  • Use the git drift check (below) to verify your file-ledger is still accurate before trusting it in a retry.
  • Context hygiene beats volume. A working-memory file that forces you to distil is more valuable than pasting raw output. Keep the live context window lean — prune irrelevant history and long raw outputs; summarize; send only high-signal tokens. Context rot (window stuffed with low-signal content) degrades a cheap model's attention faster than token-count alone suggests.

Read the full file on GitHub · 71 lines

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 · 71 lines · 26 tokens per session scan A 33db1a6ee310

Subscribe to this mod's changes

working-memory is a cursor rule published in the GitHub repository AG-Dmitry/composer-enhanced (3 stars, last pushed 2mo ago), licensed MIT. It adds 26 tokens to every session and 662 once invoked, about $0.0001 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.