Best MCP servers for codebase memory with Claude Code, compared

“Codebase memory” is four different jobs wearing one name. Before picking an MCP server for Claude Code, decide which memory you’re actually missing — most disappointment with these tools is a category error, not a quality problem. Here’s the honest map, including where Selvedge (that’s us) is not the right pick.

The four jobs

Job	The question it answers	Representative servers
Code intelligence	”How is this codebase structured — who calls what?“	codebase-memory-mcp, OpenLore
Session context	”What was I doing before this session died?”	Memory Keeper
General memory	”What do we know about this user/project?“	mem0, official server-memory, claude-code-memory
Decision provenance	”Why is the code like this — and what was already tried?”	Selvedge

The servers, side by side

Server	What it remembers	Storage	Setup	License
codebase-memory-mcp	A knowledge graph of code structure — functions, classes, calls, imports — via 15 analysis tools (search, trace paths, dead code, architecture)	SQLite-backed graph in ~/.cache/	Single static binary, one-line installer	MIT
OpenLore	A derived map of code state — tree-sitter call graph across 18 languages + 12 IaC ecosystems — plus commit-gated decision notes	SQLite graph in .openlore/	One-time repo index; 73 MCP tools across three presets (13 in the default)	MIT
Memory Keeper	Session context — categorized items (task/decision/progress/note), checkpoints, file caching, git-branch channels, ~38 tools	SQLite (WAL) in ~/mcp-data/	Node: claude mcp add memory-keeper npx mcp-memory-keeper	MIT
mem0	Facts and preferences as vector memory, retrieved by semantic similarity	Vector store; self-hosted or hosted API	Python/API; embeddings required	Apache-2.0
Official server-memory	A manually-built knowledge graph of entities, relations, observations	Local JSON file	npx one-liner	MIT
claude-code-memory	A Neo4j-backed knowledge graph of coding activity and relationships	Neo4j (runs separately)	Requires a Neo4j instance	—
Selvedge	Per-entity change history with the agent’s own reasoning, outcomes, expiry — written live as the agent works	SQLite under .selvedge/ next to your code	pip install selvedge && selvedge setup, or the Claude Code plugin	MIT

Which one to pick

Pick codebase-memory-mcp if agents keep misunderstanding your architecture — it’s the strongest structure index here, and its dead-code and call-tracing tools are things Selvedge doesn’t do at all.

Pick OpenLore if you want a structure index and decision notes from one deterministic, local-first server — it’s the closest thing to Selvedge in the category, and we keep a full comparison. The difference that decides it: OpenLore derives its memory from code state and purges rejected decisions from the queryable store after each sync (a decision that was synced before rejection leaves an annotation in the spec markdown — the queryable record is what disappears). Selvedge keeps rejected paths queryable.

Pick Memory Keeper if your pain is losing your place — long sessions dying with un-checkpointed context. That’s its lane.

Pick mem0 or the official memory server if you want durable facts and preferences that follow a user across applications.

Pick Selvedge if the pain is “why is the code like this?” and “the agent just re-made a mistake we reverted in March.” It’s the only server on this list where a rejected path stays queryable, reasoning attached — which is exactly the record prior_attempts needs to stop the agent before the edit. That’s the lane we built for: decision provenance, with a queryable audit trail and Agent Trace interop.

And they compose. These aren’t competitors for one slot — Claude Code runs multiple MCP servers happily. A structure index plus a decision log is a strictly better-informed agent than either alone. Besides OpenLore’s decision notes (covered above), one overlap to watch: Memory Keeper’s “decision” category items are freeform session notes, while Selvedge decisions are entity-keyed events with outcomes and expiry — if you want decisions the next session can act on automatically, that difference is the whole game.

What Selvedge is

The decision-provenance row of that table, and nothing else on it: captured testimony — what the agent said at change time — rather than an index derived from code state. 8 MCP tools that also work as a plain CLI: log_change, prior_attempts, blame, diff, history, changeset, search, stale_decisions. SQLite next to your code, MIT licensed, no LLM in the core.

pip install selvedge
selvedge setup   # detects Claude Code / Cursor / Copilot, installs the prompt block

Full quickstart →

FAQ

What is the best MCP server for codebase memory with Claude Code?

It depends which memory you’re missing. For code structure and navigation, codebase-memory-mcp or OpenLore. For resuming interrupted sessions, Memory Keeper. For user preferences across apps, mem0 or the official memory server. For decision history — why changes were made and what was already tried and rejected — Selvedge.

Do these memory servers replace each other?

Mostly no — they remember different things. A structure index, a session checkpoint, and a decision log answer different questions, and Claude Code can run several MCP servers side by side.

Which codebase memory servers work fully offline?

codebase-memory-mcp, OpenLore, Memory Keeper, the official memory server, and Selvedge all store locally (SQLite or JSON) with no required cloud. mem0 can self-host; its hosted API is optional.

What does Selvedge remember that the others don’t?

Rejected paths, with the reasoning. Selvedge events carry the agent’s own reasoning per entity, plus outcomes — and a rejected decision stays queryable instead of being dropped (OpenLore purges rejected decisions from its queryable store after each sync). That’s what lets prior_attempts tell an agent the change it’s about to make was already tried and reverted, and why.