Hydrate vs ClawMem
ClawMem is the tool most architecturally similar to Hydrate — same 90/10 hook model, same local-SQLite-first design. It has the best retrieval stack of any personal-memory tool we've teardown'd. Hydrate's edge is everything above a single developer: team canon, cross-vendor, orchestration.
Most comparisons on this site contrast two tools that chose opposite philosophies. ClawMem is the exception: it made most of the same calls Hydrate did. Both are local-SQLite-first, both run on the "90/10" principle (memory arrives automatically via hooks, so the developer almost never queries it), and both hook the same Claude Code surface. The honest read is that ClawMem is ahead on retrieval quality and Hydrate is ahead on the team, cross-vendor, and orchestration layers — and Hydrate is actively adopting ClawMem's retrieval mechanisms (see below).
At a glance
| Dimension | Hydrate | ClawMem |
|---|---|---|
| Language / runtime | Go stdlib | TypeScript / Bun |
| Runtime deps | None | Bun runtime |
| Storage | SQLite (local-first) | SQLite (local-first) |
| Hook / injection model | 90/10 automatic pre-prompt | 90/10 automatic pre-prompt |
| Capture | Stop hook → deterministic local distill | Claude Code hooks |
| Retrieval ranking | Similarity/recency cascade (RRF + rerank in progress) | RRF + cross-encoder rerank + co-activation |
| Recency decay | Global τ (per-type table in progress) | Per-content-type half-lives + access-extended |
| Intent routing | - | ✓ MAGMA-style |
| Memory taxonomy | Enumerated fact / canon + orchestration types | 9-type YAML taxonomy |
| Three-layer bootstrap | ✓ intent + history + live | - |
| Team canon propagation | team push/pull | - |
| Hydration packs (<code>.hpack</code>) | ✓ | - |
| Orchestration (Design / Develop) | ✓ + cross-vendor | - |
| Cross-vendor live session | ✓ Claude ↔ Codex (+ Copilot, Vibe-fork) | - Claude-only |
| Licence | Closed source (v0.2.0 beta) | MIT |
| Version referenced | v0.2.0 | v0.10.7 |
✓ present · - not present · bold + tinted cells mark the side that ships the capability.
Three-way: Hydrate vs ClawMem vs claude-mem
The three closest personal-memory tools for Claude Code, on the dimensions that actually separate them. ClawMem and Hydrate share a local, zero-cost, automatic-injection design; claude-mem takes the opposite call on compression (LLM-powered, API-metered) and injection (on-demand).
| Dimension | Hydrate | ClawMem | claude-mem |
|---|---|---|---|
| Stack | Single Go binary | TypeScript / Bun | Node + Bun + uv + worker |
| Compression | Deterministic local | Local | LLM-powered (Agent SDK) |
| Compression cost | Zero | Zero | API tokens per tool call |
| Injection | Automatic pre-prompt | Automatic pre-prompt | On-demand (MCP query) |
| Retrieval ranking | Cascade (rerank in progress) | RRF + cross-encoder rerank | Worker-service capture |
| Team canon propagation | ✓ | - | - (Pro: device sync) |
| Cross-vendor live session | ✓ | - | - |
| Orchestration | ✓ | - | - |
| Licence | Closed (v0.2.0 beta) | MIT | Apache 2.0 |
The pattern is the same one the broader landscape shows: the commodity rows are shared, and the bottom three rows — team propagation, cross-vendor live session, orchestration — are Hydrate-only. A dedicated Hydrate vs claude-mem page covers that pairing with measured benchmarks.
What is ClawMem?
ClawMem is a TypeScript/Bun, MIT-licensed memory plugin for Claude Code, architecturally the closest tool to Hydrate we've analysed (teardown verified against cloned source at v0.10.7). It shares Hydrate's core bets: local-SQLite-first storage, a "90/10" hook model where memory arrives automatically rather than on request, and the same Claude Code hook surface.
- Its genuine lead is retrieval. Reciprocal rank fusion plus cross-encoder reranking, per-content-type recency half-lives with access-extended decay, co-activation reinforcement, and MAGMA-style intent routing — the most sophisticated recall stack of any personal-memory tool in the category.
- 9-type YAML memory taxonomy classifying facts, decisions, progress, and more.
- Licence: MIT — source-available and forkable.
- Scope: single-developer, single-vendor (Claude Code). No team-propagation or cross-vendor mechanism documented.
Where ClawMem is stronger
- Retrieval ranking. ClawMem's RRF + cross-encoder rerank + co-activation pipeline is genuinely ahead of Hydrate's current similarity/recency cascade. This is real, and we say so: Hydrate has a signed-off plan to adopt the mechanisms that fit its invariants (dependency-free, fail-open, flag-gated).
- Decay modelling. Per-content-type half-lives with access-extended decay are more nuanced than Hydrate's single global decay constant (a per-type table is in progress).
- Intent routing. MAGMA-style routing of a query to the right memory lane; Hydrate doesn't have a direct equivalent yet.
- MIT licence. Source-available and forkable. Hydrate v0.2.0 is binary-only beta — if you need source for audit or compliance, ClawMem is the right answer.
Where Hydrate is stronger
- Team canon propagation.
hydrate team push/pullrounds facts through a git remote the team controls, preserving attribution into a colleague's local store. ClawMem memory is per-developer with no propagation mechanism. - Cross-vendor live-session memory. A fact written by one vendor's agent is available, in real time, to a different vendor's agent — proven Claude Code ↔ Codex on stock runtimes, plus Copilot and the Vibe fork. ClawMem is Claude-only.
- Orchestration. Hydrate's Design/Develop modes coordinate multi-agent work across vendors on the same shared-state substrate. ClawMem has no orchestration layer.
- Three-layer bootstrap. Hydrate ingests intent (
CLAUDE.md+ docs), enacted history (transcripts), and live capture as peer inputs, so a fresh install on a mature project starts with months of context. ClawMem captures from the point of install forward. - Zero runtime dependencies. One Go binary and a SQLite file — no Bun, no Node. The install and bug surface are smaller.
- Hydration packs.
.hpackarchives make onboarding a one-command, committable artefact. No ClawMem equivalent.
When to pick each one
| Scenario | Better choice | Why |
|---|---|---|
| You want the most sophisticated single-developer recall today | ClawMem | RRF + cross-encoder rerank + co-activation is ahead, now |
| You need a source-available (MIT) tool | ClawMem | Hydrate v0.2.0 is closed beta |
| Your decisions need to reach the rest of your team | Hydrate | ClawMem has no canon propagation |
| You work across Claude Code, Codex, Copilot or Vibe | Hydrate | ClawMem is Claude-only; Hydrate shares memory cross-vendor |
| You orchestrate multiple agents | Hydrate | Design/Develop modes; ClawMem has no orchestration |
| You want zero runtime dependencies | Hydrate | One Go binary vs a Bun runtime |
| You're adopting on a mature, pre-existing project | Hydrate | Three-layer bootstrap ingests existing intent + history |
No head-to-head benchmark yet — and we're adopting their ideas
Unlike the claude-mem comparison, we have not run a measured head-to-head against ClawMem. This page is feature- and architecture-based, drawn from a source-verified teardown of v0.10.7. We won't publish benchmark numbers we haven't actually produced.
We'll also be plain about influence: ClawMem's retrieval stack is good enough that Hydrate has a signed-off plan to adopt the parts that fit its invariants — reciprocal rank fusion, optional cross-encoder reranking (remote-or-skip, never a mandatory dependency), and per-content-type decay — incrementally and behind default-off flags. Borrowing the best idea in the category and saying so is more useful than pretending we invented it.