A comparison table cataloging memory systems for AI coding agents has been published on GitHub. It contrasts 74 open-source products across 79 feature items, presented as an auditable matrix in which every judgment is backed by public sources.
Open-Source · AI Memory Comparison
74 AI Memory Systems, 79 Criteria — Lined Up Without the Marketing
A single open comparison places memory engines for coding agents side by side across seven axes — every ✅ verdict backed by evidence files from public docs and source code. "No affiliation, no marketing — just facts."
74
memory systems compared
79
feature criteria scored
Feature coverage by system
Share of the scored criteria each project meets (where reported)
Standout traits per system
Fidelis
83.2% R@1
on LongMemEval-S
Kage
96.17% R@5
verified code citations
Fullerenes
64% fewer tokens
Tree-sitter code graph, SWE-bench
Midas
30–40% fewer tokens
LLM-free ingest, source tracing
Mem0
66.9% accuracy
on 600-turn conversations
Why structured memory
Context windows fill up and recall is lost between long, multi-session coding tasks. Entity extraction, conflict detection, temporal handling and code-graph integration aim to give agents persistent memory where plain RAG falls short.
The complexity trap
Heavy extraction pipelines can amplify AI sycophancy by up to 25× . Simpler, summary-based memory may strike a better real-world balance — more machinery is not always better.
The seven axes
Data Model
Search & Retrieval
Knowledge Lifecycle
Extraction Pipeline
Platform Support
Architecture
Benchmarks
Updated live — system counts and figures may shift on the public table. Built for and adaptable to coding agents like Claude Code, Codex and OpenCode.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…