Hydrate vs claude-mem

Delta: +6.29. Hydrate's worst cell (18 / 20) beat claude-mem's best (17 / 20) on scenario-lquery (Sonnet 4.6, fresh Docker containers per cell). Methodology →

claude-mem and Hydrate are the two closest tools in the session-memory space — they both hook into Claude Code to capture what happens in a session and inject it back later. Then they choose opposite architectural philosophies, and the benchmarks tell you which one works better at the canonical "did we keep the right context across the boundary?" question.

What is claude-mem?

thedotmack/claude-mem (homepage claude-mem.ai) is a Node + Bun + uv stack that captures every Claude Code tool execution via PostToolUse hooks and forwards observations to a worker service running on localhost:37777. The worker uses the Claude Agent SDK — i.e. it calls back to the Anthropic API — to compress those observations into "structured learnings" (facts, decisions, insights) extracted via XML parsing.

• Ships as a Claude Code plugin (and a Gemini CLI / OpenCode variant).
• Stack: Node.js 18+, TypeScript, Bun, uv (Python); SQLite plus ChromaDB for storage.
• Licence: Apache 2.0. A paid claude-mem Pro tier adds cross-device sync.
• Maturity: roughly 5.8k GitHub stars; maintained by Alex Newman (@thedotmack). Latest release v13.2.0 (2026-05-12) added native Windows support and first-class Gemini CLI integration.

Both tools share the same broad goal — persistent memory across Claude Code sessions. The rest of this page is about the implementation differences and what they cost.

How to install claude-mem

npx claude-mem install

The installer pulls Node 18+, Bun, and uv (Python package manager) as runtime dependencies. The worker process binds to localhost:37777 and the ChromaDB-backed memory store lives in ~/.claude-mem/. Restart Claude Code after install; the SessionStart hook fires on every new session and replays compressed summaries on the startup|clear|compact matcher.

Native Windows is supported in v13.2.0 per the README. Node 18+ is a hard requirement; the global npm install -g claude-mem ships the SDK only and is explicitly not the install path.

Architecture differences (the deep version)

Capture point

• claude-mem captures via the PostToolUse hook — fires after every individual tool execution (every Read, every Bash, every Edit). The observation is queued to the worker for AI-powered compression.
• Hydrate captures via the Stop hook — fires once when the session ends. The hook reads the canonical Claude Code JSONL transcript and runs a deterministic local distill.

Compression

• claude-mem uses the Claude Agent SDK. Each observation triggers an LLM call to extract structured XML. This costs real Anthropic API tokens — claude-mem's own README acknowledges "tokens consumed per tool execution." Over a typical session that's hundreds of background API calls.
• Hydrate uses TF-IDF plus sentence scoring plus per-clause information-density scoring. Pure Go, microsecond-level latency, zero API calls. Over a 26-hour orchestration sprint, Hydrate compressed 25.5M raw transcript tokens to 142K stored summary (0.6% retention) at $0 of compression cost.

Injection

• claude-mem injection is on-demand — Claude has to actively call an MCP search tool to retrieve memory. The model has to decide to remember to look things up.
• Hydrate injection is automatic — the UserPromptSubmit hook injects ranked context before Claude reads the prompt. The model doesn't have to know to remember; the context is just there.

Compaction survival

• claude-mem restores at SessionStart only, which means it relies on Claude Code's own (lossy) compaction summary as the input.
• Hydrate snapshots via PreCompact + SessionStart — the PreCompact hook writes the transcript tail plus a distill before Claude Code's compaction event mutates the conversation. The next session's SessionStart hook reads that snapshot first.

Infrastructure

• claude-mem needs Node 18+, Bun, uv, ChromaDB, plus a worker process on :37777.
• Hydrate is one Go binary plus a SQLite file. No runtime, no worker service, no Python.

Where claude-mem is stronger

• Granular observation capture. claude-mem can answer "what files did you read in turn 47?" because it captured each Read individually. Hydrate has the session as a whole; per-turn metadata requires re-parsing the JSONL.
• First-class Gemini CLI integration. v13.2.0 ships --ide gemini-cli as a documented install path. Hydrate reaches Gemini CLI via the generic hydrate-mcp server; claude-mem markets the integration explicitly.
• Apache 2.0 licence. Auditable, forkable. Hydrate v0.2.0 is binary-only beta. If you need source-available for compliance or audit, claude-mem is the right answer.

Where Hydrate is stronger

• Non-zero recovery on compact-survival. 27% vs 0%, n=30. Hydrate is the only memory tool that recovered in-flight tasks across the Stop → SessionStart boundary on the canonical scenario. claude-mem's failure mode is confabulation (28 of 30 cells wrong, answering with confidence); Hydrate's is restraint (8 missing, 6 asking — it returns nothing when it isn't sure). Run it yourself.
• Zero LLM cost to operate. No API call in the compression pipeline. claude-mem's PostToolUse-based capture is paying compression costs on every tool execution; over a long session that's hundreds of API calls.
• PreCompact hook intercepts Claude Code's compaction. Hydrate writes its own snapshot before Claude Code's lossy summary replaces the transcript. claude-mem reads the post-compaction state.
• No worker, no Bun, no Python. One Go binary, one SQLite file. The install surface and the bug surface are an order of magnitude smaller.
• Team sync plus hydration packs. hydrate team push/pull and .hpack archives propagate canon across developers. claude-mem has no equivalent — memory is per-developer.
• Cross-vendor bidirectional sync proven. Claude Code ↔ Mistral Vibe round-trip in the same session sequence. claude-mem claims MCP compatibility but doesn't demonstrate bidirectional round-trips.

When to pick each one

Methodology

scenario-lquery: a long-query memory-recall benchmark. The agent is given a multi-clause question that requires assembling four pieces of project context to answer correctly. Each cell runs in a fresh Docker container with the memory tool pre-loaded with the relevant facts. Scoring is 0–20 based on correctness of the four sub-claims (5 points each) with partial credit for direction-of-correctness.

compact-survival: Stop → SessionStart in-flight task recovery. Phase 1 of each cell establishes an in-flight task in one session; Phase 2 opens a fresh container and asks the agent to name the task key and its next step without the operator prompting. "Recovered" means both were named correctly without asking what they were doing.

Raw data and reproduction commands: bench/scenario-lquery/RESULTS-vs-claude-mem.md (commit 7105ede) · bench/compact-survival/RESULTS.md.