NexusMem Review 2026: Local Memory for Coding Agents That Actually Remembers What Failed

Most coding agents can read git log, but they never saw the shell command that failed with exit code 1, or the diff you abandoned halfway through. NexusMem closes that gap with a local, zero-cloud memory layer built specifically for coding agents.

What is NexusMem?

NexusMem is a local memory engine for coding agents. It normalizes shell history (with exit codes and timestamps once you install the hook), per-file git history (patches), and project docs into a single SQLite database of MemoryNodes, then serves token-budgeted, ranked context snippets — over an MCP server or a CLI. There’s no account, no cloud sync, and no telemetry; everything lives in ~/.nexusmem/ on your machine.

Key features

• Records shell commands with exit codes, git history (per-file patches), and project docs into a local SQLite DB
• Hybrid retrieval: BM25 (FTS5) plus optional vector search (sqlite-vec with Ollama) fused via Reciprocal Rank Fusion
• Token-budget packing returns ranked, pruned context chunks without calling a model to summarize
• MCP server exposes search_memory / sync_project / get_status for agent integration
• Cross-project queries over all initialized repos; content-addressed nodes (sha256) avoid duplicate ingestion

Who should use it?

NexusMem suits developers who run coding agents (Claude Code, Codex, Aider, and friends) on private or offline codebases and want the agent to recall prior attempts, failed commands, and relevant diffs without re-paying token costs for full git log -p dumps. The README cites 94–100% token savings versus dumping history wholesale.

Pros and cons

Pros: fully local and private; a thoughtful retrieval design (RRF fusion, signal/recency priors); MCP-ready; content-addressed so re-syncs are cheap.

Cons: needs Node 22+ (Node 20 lacks prebuilt better-sqlite3); semantic search requires a local Ollama instance; a young project (v0.3.1) with a smaller ecosystem than mature memory layers.

Pricing

Free and open source under the MIT license.

FAQ

Does it send my code anywhere? No — all data stays in a local SQLite DB; there’s no account or telemetry.

How does it integrate with my agent? Via the bundled MCP server (search_memory, sync_project, get_status) or the CLI.

Is vector search required? No — BM25/FTS5 works out of the box; vector search is optional and needs Ollama.