neuron vs. Claude Code auto memory, Mem0, Zep, claude-mem, agentmemory, and Beads

Where neuron sits

Neuron is scoped to one job: giving a coding agent (Claude Code, Codex CLI, Cursor, GitHub Copilot CLI) persistent, schema-enforced memory that lives in your repo as plain markdown, with a disposable SQLite index underneath for hybrid search. That’s narrower than a general-purpose agent-memory platform and different in kind from a graph-backed issue tracker — the comparisons below focus on the tools closest to that specific scope, not every memory-adjacent product.

Harness-native memory (Claude Code auto memory, Cursor memories)

The closest thing to neuron most users already have. Claude Code’s auto memory lets Claude record learnings on its own into a per-project directory under your home folder (~/.claude/projects/<repo>/memory/), with a MEMORY.md index whose first 200 lines (or 25 KB) are loaded at the start of every session and topic files read on demand. Cursor’s memories work on a similar per-user basis.

The architectural difference is scope, not quality: that memory belongs to one user on one machine and one harness, and nothing about it passes through a pull request. Neuron’s store is committed to the repo, so it is shared by every clone and every teammate’s agent regardless of which harness they use, changes to it are reviewed as diffs, and a schema declared in neuron.yaml constrains what can be written. Recall is also different in kind — a hook-driven, relevance-gated lookup per prompt rather than a fixed prefix loaded once — though for a single developer on a single tool the native feature may be all you need. The two can run side by side; nothing in neuron disables or depends on the harness’s own memory.

Mem0

Mem0 self-describes as a “Universal memory layer for AI Agents” on its GitHub repository and is cloud-hosted with a Python/LangChain-first SDK; its own blog markets a 5-minute Claude Code integration (mem0.ai/blog/claude-code-memory). Neuron’s difference is architectural, not a claim about which is better: Mem0 stores memory behind an API you call; neuron stores memory as .neuron/*.md files you can open, diff, and edit directly, with no API key or account required to run it.

Zep

Zep positions itself as “Agent memory at enterprise scale” and owns a dedicated definitional page for “temporal knowledge graph” (getzep.com/ai-agents/temporal-knowledge-graph) — an enterprise, conversational-agent framing, not a coding-agent-specific one. Neuron doesn’t use a knowledge graph at all; its schema is flat, declared per category in neuron.yaml, which is a different tradeoff (simpler mental model, no relationship-graph queries) rather than a strictly smaller one.

claude-mem

claude-mem captures Claude Code session activity, “compresses it with AI, and injects relevant context back into future sessions.” The compression step is the key architectural difference from neuron: claude-mem’s stored memory is an AI-generated summary, while neuron’s markdown files are the literal source of truth an agent (or a human) wrote — nothing summarizes or rewrites an entry on the way in.

agentmemory

agentmemory is local-only, like neuron, and supports multiple harnesses including Claude Code, Copilot CLI, and Cursor. It doesn’t declare a schema per category, enforce required fields at write time, or generate a codebase architecture blueprint — neuron’s closest local-only competitor, differentiated on the schema-enforcement and blueprint-scan mechanisms rather than on “runs locally,” which agentmemory already does too.

Beads and Beans

Beads is the closest analog to neuron’s issue-tracker category specifically: a “Distributed graph issue tracker for AI agents, powered by Dolt,” also marketed as “a memory upgrade for your coding agent.” It’s graph- and Dolt-backed rather than plain-markdown, with its own versioning model distinct from a plain-text git diff. Beans is closer to neuron’s plain-markdown-file model — a flat-file .beans directory readable by humans and machines — but is scoped to issue tracking only, with no learnings/decisions/architecture categories or relevance-gated recall.

Questions

Is neuron a replacement for Mem0 or Zep? Only if you’re specifically memory-augmenting a coding agent and want that memory to live in your repo as plain markdown. Mem0 and Zep are cloud-hosted, general-purpose agent memory layers built for a broader range of applications than coding agents alone; neuron is scoped narrowly to the coding-agent workflow and runs 100% offline.

Does neuron replace Claude Code’s built-in auto memory? They can coexist, but they solve different problems. Claude Code’s auto memory is stored per user under ~/.claude/projects/<repo>/memory/ and loaded only for you; neuron’s store is committed to the repo, so every teammate’s agent — on Claude Code, Codex, Cursor or Copilot — reads the same entries, and changes to it go through code review.

Can I use neuron alongside an existing CLAUDE.md or AGENTS.md file? Yes — they solve different problems. CLAUDE.md/AGENTS.md is static prose reread in full every prompt; neuron is schema-enforced structured entries surfaced only when relevant. Nothing about neuron requires deleting an existing CLAUDE.md/AGENTS.md file, though its content often migrates naturally into neuron’s declared-field categories over time.

Does neuron do anything graph-based, like Zep or Beads? No. Neuron’s data model is flat, schema-validated markdown entries with a disposable SQLite search index — not a knowledge graph (Zep) or a Dolt-backed dependency graph (Beads). If you need graph-native memory or multi-agent merge semantics, those tools solve for that in a way neuron doesn’t.

Limitations

The descriptions above are drawn from each project’s own public README, docs, or marketing copy as of September 2026, not independent hands-on testing of their live behavior — treat them as directional, not as neuron’s verified benchmark against a competitor, and check the linked source directly before relying on a specific claim. Tools in this space change quickly; a comparison page is inherently a snapshot, not a live sync.