Recoleta
GitHub
Source note

PROJECTMEM: A Local-First, Event-Sourced Memory and Judgment Layer for AI Coding Agents

Authors
Ripon Chandra Malo, Tong Qiu
Source
arXiv
Published
Collection
Software Intelligence
Code IntelligenceAI Coding AgentsProject MemoryMCPEvent Sourcing
Open arXivMarkdown source

Summary

PROJECTMEM adds persistent project memory and a deterministic pre-action warning layer to AI coding agents. It targets repeated failed fixes and repeated context reconstruction across coding sessions.

Problem

  • AI coding agents lose project-specific state between sessions, so they re-read files, re-derive decisions, and may try fixes that already failed.
  • The paper estimates that rebuilding project context can cost 5,000-20,000 tokens per session, which raises cost and wastes time.
  • Repeated failed fixes matter because they can send an agent through the same debugging path without new evidence or tests.

Approach

  • PROJECTMEM records development history as an append-only plain-text event log with typed events: issue, attempt, fix, decision, and note.
  • It rebuilds AI-readable summaries deterministically from the log, so the agent reads compact project memory without vector search, embeddings, or LLM fact extraction.
  • It exposes memory through MCP with read and write tools, plus CLI commands and a Markdown bridge for tools without MCP support.
  • Its precheck_file(path) gate checks the log before an edit and warns about prior failed attempts, open issues, or high-churn files tied to that path.
  • It runs locally with no telemetry or cloud dependency, redacts secrets before writing events, and can promote library-level gotchas into a machine-wide local store.

Results

  • The released implementation is a Python package with 3 runtime dependencies, a footprint under 5 MB, 14 MCP tools, 19 CLI commands, and 37 automated tests.
  • The paper evaluates usage through a 2-month self-study across 10 projects with 207 logged events.
  • The system reports estimated savings through pjm score, including hours, tokens, and dollars, but the excerpt does not give measured savings values.
  • The paper claims context rebuilding can cost 5,000-20,000 tokens per session, while memory reads have a more fixed cost after events are written.
  • The capability table compares PROJECTMEM with 13 named systems and claims it is the only listed system with all six target properties: local-first, plain-text without vector DB, event-sourced immutable storage, pre-action judgment, MCP-native access, and cross-project memory.
  • The paper does not report a controlled benchmark for coding task success rate, bug-fix accuracy, or head-to-head agent performance.

Link