Short answer. For most people Claude Code's built-in memory is the right starting point: it leads on equal weights (3.86) because it costs nothing, keeps data on your machine and needs no setup. Our own plugin, DS-ReCall, is second on equal weights (3.71). When audit and privacy come first it ranks 1st (4.05 against 3.95 for the built-in memory, a gap well inside the noise of judgment scores); for a hands-off solo developer it ranks 6th. claude-mem is the most adopted add-on (101,684 GitHub stars) and the most automatic, agentmemory gives you search without a cloud account, and Mem0 and Supermemory send session content to a hosted service.
Disclosure. Published by Devsoy, the author of DS-ReCall, which is one of the tools compared. Scores are our judgment, not measurements. Every input is listed so you can re-score.
1. What we compared, and how we chose
We looked at memory layers that carry knowledge across Claude Code sessions. The selection rule: the most adopted projects that connect to Claude Code for cross-session memory, ranked by the GitHub stars of the project (plus package downloads where available), and the built-in memory every Claude Code user already has. That gives five options: claude-mem, Mem0, Supermemory, agentmemory and Claude Code's built-in memory. We added our own DS-ReCall as the sixth.
We measured on 19 October 2026 from the GitHub API, npm and PyPI download counts, each project's own documentation and the Claude Code documentation (version 2.1.299). Vendors' own benchmark claims are reported as claims; we did not re-run them.
One fact limits every popularity number: stars measure attention, not users, and download counts belong to SDK or library packages, not to the Claude Code plugins. The only measured install count we found is the Remember plugin's page in Anthropic's directory (51,442 installs on 19 October 2026). That plugin is outside the main comparison; see section 9.
2. The six at a glance
| Tool | What it is | GitHub stars | Contributors | Licence | Latest release |
|---|---|---|---|---|---|
| Claude Code built-in memory | CLAUDE.md + auto memory, ships with Claude Code | — | — | part of Claude Code | Claude Code 2.1.299 |
| DS-ReCall | Curation plugin for CLAUDE.md and rules, no database | 0 | 1 | MIT | v0.3.1 (2026-10-19) |
| agentmemory | Local memory server for many agents | 29,667 | 51 | Apache-2.0 | v0.9.30 (2026-10-06) |
| claude-mem | Plugin + local worker service | 101,684 | 290 | Apache-2.0 | v14.0.1 (2026-10-17) |
| Mem0 (Claude Code plugin) | Cloud memory platform + plugin | 67.4K (whole project) | 410 | Apache-2.0 | plugin 0.3.1 · SDK 2026-09-25 |
| Supermemory (Claude Code plugin) | Hosted memory API + plugin | 31.3K (main project) · 2.8K (plugin repo) | 16 | MIT per README; no LICENSE file | plugin pushed 2026-10-19 |
Swipe the table sideways to see every column.
3. How each one remembers
| Tool | Capture | Storage | Recall | Model calls |
|---|---|---|---|---|
| Claude Code built-in memory | Claude writes auto-memory notes itself; you write CLAUDE.md | Plain Markdown on your machine (~/.claude/projects/<project>/memory/ and CLAUDE.md) | First 200 lines or 25 KB of MEMORY.md loaded each session; topic files read on demand | No separate call: it runs inside your session |
| DS-ReCall | Hooks open a stub and back up each session; consolidation is a deliberate /wrap step, so you control what is remembered | Plain Markdown in your own repository (CLAUDE.md, rules, docs); no database | None of its own: the curated index loads natively; older history is searched with grep | None from the hooks; /wrap runs as a normal prompt in your session |
| agentmemory | Automatic: 12 hooks | Local server with its own engine; state on disk under your home folder | Keyword (BM25) by default; vectors and graph when embeddings are on; 54 MCP tools | Optional: keyless mode needs no LLM; compression by a model only if you configure one |
| claude-mem | Automatic: 6 lifecycle hooks | Local SQLite + Chroma vectors behind a Bun worker; optional cloud sync | Keyword + semantic search; 3 MCP tools reveal detail step by step | Yes: an observer model writes the summaries (hosted trial, your own key, or your Anthropic plan) |
| Mem0 (Claude Code plugin) | Automatic: hooks batch messages, answers and changed files | Mem0 Platform (cloud); account and API key required | Semantic + keyword + entity matching; top 5 memories injected on the first prompt; search tool | Yes: extraction runs on Mem0's service |
| Supermemory (Claude Code plugin) | Automatic: saves the conversation at session end (Stop hook) | Supermemory cloud; API key or browser login | A hook searches each substantive prompt and injects fresh matches; hosted MCP tools | Yes: processing happens on the vendor's side |
Swipe the table sideways to see every column.
4. What leaves your machine
This is the axis where the tools differ most, and the one most comparisons skip.
- Claude Code built-in memory: Local files, machine-local, editable
- DS-ReCall: Nothing leaves your machine: the hooks make no network requests
- agentmemory: Local; a privacy filter strips secrets before storing
- claude-mem: Local database; summarising runs through the provider you pick
- Mem0 (Claude Code plugin): Sent after credential redaction: messages, answers, changed files, short test results
- Supermemory (Claude Code plugin): New conversation content is sent; an offline filter masks common credential patterns but cannot catch every secret (their wording)
5. How we scored
Seven axes, each scored 1 to 5. A higher number is better on every axis; for cost and dependencies, 5 means nothing extra to carry.
- A1 · Hands-off capture. How much remembering happens without you doing anything.
- A2 · Recall power. Finding and injecting the right past knowledge as history grows (search, ranking).
- A3 · Auditability. Can you read, correct and validate what is remembered?
- A4 · Data control. How much conversation content stays on your machine; whether an account is needed.
- A5 · Cost and dependencies. Extra model calls, accounts, services and runtimes you must carry (5 = nothing extra).
- A6 · Setup and operations. Install steps, background processes, platform friction.
- A7 · Multi-session and team. Parallel sessions, shared memory across people or agents.
Three weightings show how the ranking moves with what you value: equal weights; a hands-off solo developer (capture, recall and setup count most); and audit and privacy first (auditability, data control and dependencies count most).
Score matrix
| Tool | A1 | A2 | A3 | A4 | A5 | A6 | A7 | Equal-weight average |
|---|---|---|---|---|---|---|---|---|
| Claude Code built-in memory | 4 | 2 | 4 | 5 | 5 | 5 | 2 | 3.86 |
| DS-ReCall | 2 | 2 | 5 | 5 | 5 | 3 | 4 | 3.71 |
| agentmemory | 5 | 4 | 3 | 4 | 3 | 2 | 4 | 3.57 |
| claude-mem | 5 | 4 | 3 | 3 | 3 | 3 | 3 | 3.43 |
| Mem0 (Claude Code plugin) | 4 | 4 | 3 | 2 | 2 | 3 | 4 | 3.14 |
| Supermemory (Claude Code plugin) | 4 | 4 | 2 | 2 | 2 | 4 | 4 | 3.14 |
Swipe the table sideways to see every column.
A1 Hands-off capture · A2 Recall power · A3 Auditability · A4 Data control · A5 Cost and dependencies · A6 Setup and operations · A7 Multi-session and team
Ranking under each weighting
Equal weights
- Claude Code built-in memory · 3.86
- DS-ReCall · 3.71
- agentmemory · 3.57
- claude-mem · 3.43
- Mem0 (Claude Code plugin) · 3.14
- Supermemory (Claude Code plugin) · 3.14
Hands-off solo developer
- Claude Code built-in memory · 3.75
- claude-mem · 3.75
- agentmemory · 3.69
- Supermemory (Claude Code plugin) · 3.50
- Mem0 (Claude Code plugin) · 3.38
- DS-ReCall · 3.06
Audit and privacy first
- DS-ReCall · 4.05
- Claude Code built-in memory · 3.95
- agentmemory · 3.55
- claude-mem · 3.30
- Mem0 (Claude Code plugin) · 3.00
- Supermemory (Claude Code plugin) · 2.90
Scores are our judgment on a 1 to 5 scale, not measurements. We are the authors of DS-ReCall, and the choice of axes and weights favours what we care about, which is why we show three weightings, including one where DS-ReCall ranks last. The scoring sheet lists every score, every weight and a one-line rationale for each cell; change a weight and the ranking changes.
Maturity (adoption, contributors, release cadence, age) is not part of the score. Our first draft scored it as an eighth axis. We removed it because this report asks what each tool does and how it behaves, and maturity is a risk factor rather than a function. The facts stay visible in the overview table. With maturity scored, DS-ReCall ranked 4th on equal weights, 2nd on audit and privacy first and 6th for the hands-off developer.
6. Best for, and watch out
Claude Code built-in memory
- Best for: Most people: no setup, free, local and plain Markdown.
- Watch out: Claude decides what to save; only the start of MEMORY.md loads; no search and no validation.
DS-ReCall
- Best for: Teams and client work that must be able to say what the AI was told and what was dropped; project-scoped, controlled memory; parallel sessions; no data leaving the repository.
- Watch out: No search layer; /wrap is deliberate friction; a single maintainer, zero stars and a young codebase.
agentmemory
- Best for: Local-first users who want strong search and multi-agent support without a cloud account.
- Watch out: Heavier runtime (engine and four ports); manual engine step on native Windows; 577 open issues; benchmarks are self-run.
claude-mem
- Best for: Solo developers who want the most automatic, most widely adopted option and accept a background service.
- Watch out: Moving parts (Bun worker, vector store, sign-in flow); a model writes your summaries; savings figures are the vendor's own.
Mem0 (Claude Code plugin)
- Best for: Teams already on Mem0 who want shared project memory with personal scopes.
- Watch out: Needs a cloud account and sends session content to the service; headline benchmarks are vendor-reported.
Supermemory (Claude Code plugin)
- Best for: Developers who want hosted recall shared across Claude Code, Codex and OpenCode with little setup.
- Watch out: Cloud account; conversation content leaves the machine; the plugin repository is young and has no detected LICENSE file.
7. DS-ReCall, stated plainly
What it does well
- Manual by design: nothing enters the memory unless a person approves it at the /wrap step. That is where its control comes from: what is remembered is what you decided, per project, in files you can read and diff.
- Every fact is dated and status-tagged, one fact per file, and a validator finds broken, orphaned and unindexed notes.
- The hooks make no model calls and no network requests; there is no database and nothing to host.
- Session locks and parallel-session awareness; a commit gate that warns by default and can block.
- 140 self-test cases pass, and a negative harness proves the suite goes red when real defects are injected (34 of 34 exercisable defects caught in the 0.3.0 run, one guard declared not exercisable).
What it does not do
- No retrieval layer: recall is the curated index plus grep. This is the biggest capability gap against the tools above.
- Consolidation needs a deliberate /wrap; it is friction by design and only as good as the session.
- Young: 0 stars, one maintainer, 3.5 months old, tested mainly on one Windows workspace and one Node version.
- The commit-gate warning has not yet been observed in a live, logged-in Claude Code session by us; the tool to prove it ships with the plugin.
8. Which one should you pick?
- You just want it to work: Start with the built-in memory. Add a search tool only when you feel the gap.
- You want the most automatic recall and accept a background service: claude-mem
- You want search across agents, kept local: agentmemory (or engram, which is smaller)
- You already use a hosted memory platform and accept the cloud: Mem0 or Supermemory
- You must audit what is remembered, keep data in the repository and run parallel sessions: DS-ReCall
9. Also considered
- Remember (Anthropic directory). Listed in Anthropic's directory with 51,442 installs on its page (2026-10-19), the only measured install count we found. It summarises sessions with Claude Haiku (vendor: under $0.01 per save). Not ranked: no repository stars to compare and a different model.
- engram. 7,444 stars, MIT, Go binary with SQLite FTS5, agent-driven deliberate saves, 14 agents. The closest relative in philosophy and a strong option; smaller than the top five.
- basic-memory. 4,277 stars, AGPL-3.0, Markdown files plus MCP.
- memsearch. 2,759 stars, MIT, Markdown source with a Milvus index.
- total-recall. 202 stars; no commits since 2026-02-12.
Download the report
The full report is a 14-slide PDF, and the scoring sheet is a CSV with every score, weight and rationale. DS-ReCall is open source under the MIT licence.
Frequently asked questions
What is the most popular memory tool for Claude Code?
By GitHub stars on 19 October 2026: claude-mem (101,684), then Mem0 (67,426, whole project), Supermemory (31,298 for the main project, 2,791 for its Claude Code plugin) and agentmemory (29,667). Stars measure attention, not users. The only measured install count we found is the Remember plugin in Anthropic's directory, at 51,442 installs.
Does Claude Code have built-in memory?
Yes. Two mechanisms carry knowledge across sessions: CLAUDE.md files that you write, and an auto memory that Claude writes itself under ~/.claude/projects/<project>/memory/. The first 200 lines or 25 KB of MEMORY.md load at the start of every session. You can browse, edit or switch it off with /memory.
Which Claude Code memory tools keep my data on my machine?
The built-in memory and DS-ReCall keep everything local (DS-ReCall's hooks make no network requests). agentmemory runs a local server and lets you decide whether any external model provider is used. claude-mem keeps its database local, but its summarising runs through the provider you choose. The Mem0 and Supermemory plugins send session content to the vendor's service.
Is DS-ReCall better than claude-mem?
Not overall. claude-mem is more automatic, has real search and far wider adoption. DS-ReCall suits a different need: teams and client work that must be able to say what the AI was told and what was dropped, with parallel sessions and no data leaving the repository. It has no search layer, needs a deliberate /wrap step and is a young project with a single maintainer.
How did you score the tools?
Seven axes, scored 1 to 5 by us, with three weightings to show how the ranking moves. The scoring sheet (CSV) lists every score, every weight and a one-line rationale per cell, so you can change a weight and re-rank. We are the authors of DS-ReCall, so we say plainly that the scores are judgment, not measurement.
Follow the conversation
The announcement posts and discussion: LinkedIn · Facebook · Instagram
Related reading
Sources
- GitHub REST API: stars, forks, licence, release, contributors (2026-10-19)
- npm downloads API (last month: 2026-09-17 to 2026-10-16) and pypistats (mem0ai)
- claude-mem README (main branch)
- Mem0 README and Claude Code integration docs
- claude-supermemory README
- agentmemory README
- Claude Code docs: How Claude remembers your project
- Remember plugin page (Anthropic directory)
- engram README
- DS-ReCall repository, CHANGELOG and self-tests
All data was measured on 19 October 2026: GitHub REST API (stars, forks, licence, releases, contributors), npm downloads for 17 September to 16 October 2026, PyPI downloads for mem0ai, each project's README or documentation, and the Claude Code documentation. Scores are the authors' judgment. Download counts are for SDK or library packages, not plugin installs. Also available in Turkish.