acm

okitegami_paper

Associative Context Memory (ACM): Persistent Experience-Based Memory for LLM Coding Agents Beyond Context Window Dependency

Hiroshi Yamato — Keio University / Signal compose Inc.

---

Overview

This repository contains the research paper draft and supporting documents for ACM (Associative Context Memory), an external persistent memory architecture for LLM coding agents.

ACM captures success and failure experiences from implicit user feedback signals (interrupt events, rewind operations, corrective instructions) and injects relevant past experiences into new sessions, enabling agents to learn across sessions without relying on context window capacity.

Key Contribution

A taxonomy of implicit feedback signals in coding agent interactions, with PostToolUseFailure.is_interrupt (Claude Code hooks API) identified as the strongest available signal of user dissatisfaction.

Repository Structure

• docs/acm-preprint-draft.md — Paper draft (v0.11)
• docs/acm-preprint-draft-ja.md — Paper draft Japanese (v0.11)
• docs/SPECIFICATION.md — Implementation specification (Single Source of Truth)
• docs/ROADMAP.md — Implementation roadmap
• docs/acm-handoff.md — Research context and design decisions
• docs/session-log.md — AI collaboration process log (auto-updated by Stop hook)
• docs/pdf/ — IEEEtran format PDFs (EN/JA)
• adr/ — Architecture Decision Records
• src/ — Implementation (ACM MCP server)
• experiments/ — Task definitions, procedures, and results

Status

• Research design
• Related work review (ADR 0001)
• Phase 0 feasibility probe (ADR 0002)
• Paper draft v0.11
• [~] arXiv submission (pending endorsement)
• Implementation
• Experimental evaluation

Preprint

arXiv submission in progress (pending endorsement). Link will be added upon publication.

---

Signal compose Inc. / Keio University