robo-cortex: Git-Anchored, Evidence-Based Memory for AI Coding Agents

Large language model (LLM) coding agents increasingly rely on persistent memory to carry lessons across sessions, yet most existing memory stores treat a stored memory as trustworthy indefinitely: a note recorded about a piece of code remains "active" long after the code it described has changed. We present robo-cortex, a local-first memory store for AI coding agents built around a single principle we call knowledge integrity: a memory is only as trustworthy as the code it points at. Every memory may be anchored to the Git blob hash of the source file(s) it describes; a lazy staleness check, triggered on every read, recomputes those hashes and flags any memory whose underlying code has changed for review, while healing memories whose code has reverted to its recorded state. Memories carry explicit, mechanically-derived evidence strength and author-stated confidence rather than opaque relevance scores. The system uses no embeddings, no vector database, and no cloud service; it is pure Python standard library over a single SQLite file per repository, exposed through a command-line interface, a Python SDK, and eight Model Context Protocol (MCP) tools. We describe the design in full and report an honest empirical evaluation: six of six pre-registered functional scenarios pass; a controlled token-economy benchmark on a small subject model (Claude Haiku 4.5, n=5 per arm across three real historical bugs, 30/30 correct fixes) shows a median 11.9% reduction in fresh tokens and 15.1% reduction in API-equivalent cost when the agent is supplied a retrieved context pack, with a recorded memory breaking even after roughly five reuses. We keep every limitation of these measurements in view: a single small model, a single self-documenting codebase, three tasks, and modest sample sizes. The git-blob-hash anchoring, the code-change-triggered invalidation-and-healing lifecycle, and the evidence model are, to our knowledge, novel in combination; the underlying content-addressable primitive is Git's.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-07-26
DOI
https://doi.org/10.5281/zenodo.21610870
Primary Topic
Multimodal Machine Learning Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

robo-cortex: Git-Anchored, Evidence-Based Memory for AI Coding Agents

Sorin Neagu
Zenodo (CERN European Organization for Nuclear Research)
Multimodal Machine Learning Applications
preprint

robo-cortex: Git-Anchored, Evidence-Based Memory for AI Coding Agents

Sorin Neagu
preprint en

Abstract

Large language model (LLM) coding agents increasingly rely on persistent memory to carry lessons across sessions, yet most existing memory stores treat a stored memory as trustworthy indefinitely: a note recorded about a piece of code remains "active" long after the code it described has changed. We present robo-cortex, a local-first memory store for AI coding agents built around a single principle we call knowledge integrity: a memory is only as trustworthy as the code it points at. Every memory may be anchored to the Git blob hash of the source file(s) it describes; a lazy staleness check, triggered on every read, recomputes those hashes and flags any memory whose underlying code has changed for review, while healing memories whose code has reverted to its recorded state. Memories carry explicit, mechanically-derived evidence strength and author-stated confidence rather than opaque relevance scores. The system uses no embeddings, no vector database, and no cloud service; it is pure Python standard library over a single SQLite file per repository, exposed through a command-line interface, a Python SDK, and eight Model Context Protocol (MCP) tools. We describe the design in full and report an honest empirical evaluation: six of six pre-registered functional scenarios pass; a controlled token-economy benchmark on a small subject model (Claude Haiku 4.5, n=5 per arm across three real historical bugs, 30/30 correct fixes) shows a median 11.9% reduction in fresh tokens and 15.1% reduction in API-equivalent cost when the agent is supplied a retrieved context pack, with a recorded memory breaking even after roughly five reuses. We keep every limitation of these measurements in view: a single small model, a single self-documenting codebase, three tasks, and modest sample sizes. The git-blob-hash anchoring, the code-change-triggered invalidation-and-healing lifecycle, and the evidence model are, to our knowledge, novel in combination; the underlying content-addressable primitive is Git's.

Zenodo (CERN European Organization for Nuclear Research)
Evotec (United States) (US)
Multimodal Machine Learning Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

robo-cortex: Git-Anchored, Evidence-Based Memory for AI Coding Agents — Sorin Neagu · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS