projectmem adds a pre-action gate to stop AI agents repeating failed fixes
projectmem is an open-source, local-first memory and judgment layer for AI coding agents that records development history as an append-only event log and warns agents before they repeat previously failed actions.
Score breakdown
The pre-action gate introduces a governance layer that actively prevents AI coding agents from repeating known-failed actions, addressing a token-costly statelessness problem the authors identify as a bottleneck in current AI-assisted development.
- 01Authors: Ripon Chandra Malo and Tong Qiu; published on ArXiv on 2026-06-10.
- 02Stateless AI coding agents may consume an estimated 5,000–20,000 tokens per session reconstructing project context.
- 03projectmem records development as an append-only, plain-text event log of typed events: issues, attempts, fixes, decisions, and notes.
Ripon Chandra Malo and Tong Qiu identify a core inefficiency in current AI coding assistants: they are largely stateless, forcing each new session to re-read project files, re-derive prior decisions, and potentially repeat debugging attempts that already failed. The authors estimate this context reconstruction consumes an estimated 5,000–20,000 tokens per session, arguing the bottleneck is missing project memory rather than model capability.
To address this, they introduce projectmem, an open-source, local-first memory and judgment layer.
To address this, they introduce projectmem, an open-source, local-first memory and judgment layer. The system records development history as an append-only, plain-text event log of typed events — issues, attempts, fixes, decisions, and notes — and deterministically projects that log into compact, AI-readable summaries delivered through the Model Context Protocol (MCP). Crucially, projectmem goes beyond passive storage by adding a deterministic pre-action gate that warns an agent before it repeats a previously failed fix or edits a known-fragile file. The authors frame this design philosophy as "Memory-as-Governance": memory that actively influences the agent's next action rather than merely answering queries.
The system runs fully offline with no telemetry, and its immutable log doubles as a provenance trail for reproducible, auditable AI-assisted development. projectmem is packaged as a three-dependency Python library featuring 14 MCP tools, 19 CLI commands, and 37 automated tests. The authors evaluated it through a two-month self-study spanning 10 projects with 207 logged events. Source code is available at https://github.com/riponcm/projectmem.
Key facts
- 01Authors: Ripon Chandra Malo and Tong Qiu; published on ArXiv on 2026-06-10.
- 02Stateless AI coding agents may consume an estimated 5,000–20,000 tokens per session reconstructing project context.
- 03projectmem records development as an append-only, plain-text event log of typed events: issues, attempts, fixes, decisions, and notes.
- 04Summaries from the event log are served to agents via the Model Context Protocol (MCP).
- 05A deterministic pre-action gate warns agents before repeating a previously failed fix or editing a known-fragile file — framed as 'Memory-as-Governance'.
- 06Ships as a three-dependency Python package with 14 MCP tools, 19 CLI commands, and 37 automated tests.
- 07Evaluated via a two-month self-study across 10 projects comprising 207 logged events; runs fully offline with no telemetry.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 11, 2026 · 08:34 UTC. How this works →