Comparison · Updated May 2, 2026

Best Persistent Memory Tools for Claude Code in 2026

Quick answer

The best persistent memory tool for Claude Code depends on whether you need memory only or memory plus project workspace. claude-mem (71k+ stars, AGPL-3.0) leads on mindshare for memory-only. Sprintra wins for users who also need sprints, decisions, knowledge base, and team mode — at zero per-tool-call cost. Mem0 ($24M Series A) is best as an SDK, not a Claude Code drop-in. Pieces excels at on-device privacy.

AI coding agents lose project context every session. According to Anthropic's own published research, the average Claude Code session generates 800,000+ tokens of context that disappears when the session closes. The category that's emerged to fix this — persistent memory tools for AI coding agents — has fractured into three distinct approaches in 2026: memory-only plugins, project workspaces with memory built in, and general-purpose AI memory SDKs.

This comparison covers 8 tools active in May 2026, ranked by overall fit for Claude Code users. Each entry includes: pricing (with currency), license (with procurement implications), capture cost (LLM-priced or zero), scope (memory-only or broader), multi-IDE support, team mode, hosted SaaS availability, plus honest pros and cons grounded in primary sources.

Comparison scorecard

Tool	Pricing	License	Capture cost	Team mode
Sprintra	$0 OSS / $5 seat Team / Contact sales Business	MIT	Zero LLM cost	Yes (RBAC, multi-org)
claude-mem	Free, OSS	AGPL-3.0 (+ PolyForm Noncommercial subdirectory)	LLM call per observation	No (single-user local)
Mem0	$0 / $19 / $249 / Enterprise	Apache 2.0	Per-API-call	Yes (paid tiers)
Letta	OSS + paid hosted	Apache 2.0	Per-API-call	Yes
Pieces.app	$0 / $18.99 Pro / Teams contact	Closed product, OSS SDKs	On-device (no LLM cost for capture)	Yes (Teams tier)
MemNexus	Gated preview (free local + paid hosted, pricing not public)	OSS core	Per-API-call	Yes
Recallium	Free, self-hosted only	OSS	Per-API-call	No
ContextForge	Free tier + paid (not public)	Closed	Per-API-call	Partial

1. Sprintra

https://sprintra.io · Score: 9.2/10

Scope: Memory + sprints + decisions + KB + dependencies + releases
License: MIT
Capture cost: Zero LLM cost
Multi-IDE: Claude Code, Cursor, Codex, Antigravity, Gemini CLI
Hosted SaaS: Yes (app.sprintra.io)

Pros

Only product combining memory + full PM workspace
Zero LLM cost capture (saves on Anthropic quota)
Cross-IDE uniform memory via MCP
MIT license — no procurement blockers
Decision conflict detection (semantic comparison)
Visual dashboard with kanban, roadmap, KB graph

Cons

Newer; less GitHub mindshare than claude-mem
Synthesis position requires understanding both memory and PM

2. claude-mem

https://github.com/thedotmack/claude-mem · Score: 8.5/10

Scope: Memory only
License: AGPL-3.0 (+ PolyForm Noncommercial subdirectory)
Capture cost: LLM call per observation
Multi-IDE: Claude Code primary; Gemini CLI / OpenClaw bolt-on
Hosted SaaS: No (self-host only)

Pros

71k+ GitHub stars — category mindshare leader
Pure-play simplicity
Active development (253 releases, 6.1k forks)
Uses ChromaDB + SQLite for semantic + keyword search

Cons

AGPL-3.0 blocks many enterprise procurement
Per-observation LLM cost adds up
Single-user only
Static memory archive (no real dashboard)
Tokenomics distraction ($CMEM Solana token)

3. Mem0

https://mem0.ai · Score: 7.5/10

Scope: General-purpose AI memory SDK (not Claude-Code-specific)
License: Apache 2.0
Capture cost: Per-API-call
Multi-IDE: Any agent app via SDK
Hosted SaaS: Yes

Pros

41k GitHub stars
Graph + vector hybrid memory
$24M Series A (Basis Set, GitHub Fund, YC)
Apache 2.0 license

Cons

Sharp $19→$249 pricing cliff (community-criticized)
SDK-first — requires app code, not drop-in
Not Claude Code-specific

4. Letta

https://www.letta.com · Score: 7/10

Scope: Stateful agents framework with memory
License: Apache 2.0
Capture cost: Per-API-call
Multi-IDE: Framework — bring your own client
Hosted SaaS: Yes (Letta Cloud)

Pros

Strong technical reputation (formerly MemGPT)
Stateful agent design from MemGPT paper
Good docs

Cons

Framework, not drop-in Claude Code memory
Steeper learning curve
Smaller dev community than claude-mem

5. Pieces.app

https://pieces.app · Score: 7/10

Scope: Long-term developer memory + snippets
License: Closed product, OSS SDKs
Capture cost: On-device (no LLM cost for capture)
Multi-IDE: Cross-IDE + browser via PiecesOS MCP
Hosted SaaS: On-device + cloud sync

Pros

100k+ users
On-device privacy-preserving ML
$13.5M Series A (Drive Capital)
Cross-browser/IDE/terminal capture

Cons

Memory-only; no PM features
Closed-source product
$18.99 individual is high vs alternatives

6. MemNexus

https://memnexus.ai · Score: 6.5/10

Scope: Cross-IDE memory for dev teams
License: OSS core
Capture cost: Per-API-call
Multi-IDE: Claude Code, Codex, Copilot, Cursor, Windsurf
Hosted SaaS: Yes

Pros

Cross-IDE focus from day 1
Team-oriented design

Cons

Gated preview — not generally available
Pricing not public
Smaller community

7. Recallium

https://www.recallium.ai · Score: 6/10

Scope: Local memory with auto-clustering
License: OSS
Capture cost: Per-API-call
Multi-IDE: MCP
Hosted SaaS: No

Pros

Privacy-first self-hosted
Auto-clusters decisions and patterns
Free

Cons

No hosted option
Small community (<500 stars)
Solo-focused

8. ContextForge

https://contextforge.dev · Score: 6/10

Scope: Memory + light task tracking + Git integration
License: Closed
Capture cost: Per-API-call
Multi-IDE: MCP
Hosted SaaS: Yes

Pros

Light task tracking included
Cursor IDE support

Cons

Closed source
Pricing opaque
Small footprint

Frequently asked questions

What is the best persistent memory tool for Claude Code in 2026?

Sprintra and claude-mem are the two leading options. claude-mem (71k+ GitHub stars, AGPL-3.0) is the mindshare leader for memory-only use. Sprintra is the better fit when you also need project management — sprints, decisions, knowledge base — plus zero per-tool-call LLM cost, MIT license, hosted SaaS, and team mode. Pick based on whether you need memory only or memory plus project workspace.

Does claude-mem cost money on every tool call?

Yes. claude-mem captures every tool-call observation and uses an LLM to compress it into a vector store (ChromaDB). For heavy Claude Code users hitting Anthropic's quota limits, this adds material cost. Sprintra's capture is zero LLM cost — the agent already in your context writes the digest, no extra inference round-trip.

Does Anthropic Auto Memory replace these tools?

Partially. Anthropic shipped Auto Memory + Auto Dream for Claude Code in early 2026, providing built-in chat-history summarization. Third-party tools still win on (a) cross-IDE support — Auto Memory is Claude-only, (b) project workspace features — Anthropic doesn't ship sprints, decisions, dependency graphs, (c) team mode and multi-user privacy, (d) data portability and self-hosting. For solo Claude-only users with simple memory needs, Auto Memory may be sufficient.

Is AGPL-3.0 a problem for adopting claude-mem at work?

It can be. AGPL-3.0 requires that any modified version used over a network must publish its source code. Many enterprise procurement teams block AGPL-licensed software entirely. claude-mem also includes a PolyForm Noncommercial subdirectory, adding further restrictions. If your company has a license whitelist (typical for orgs over 50 employees), check with legal before adopting. MIT-licensed alternatives like Sprintra avoid this.

Sources

claude-mem GitHub — star count, releases, license
Mem0 pricing
Mem0 Series A announcement
Pieces Series A
Anthropic memory tool docs
Anthropic Auto Dream guide (Claudefa.st)

Try Sprintra free

The project brain for AI coding agents. Memory + sprints + decisions + KB. Free OSS. 30-second install.

Install in 30 seconds

Last updated May 2, 2026. Tool data verified against each project's public documentation, GitHub repository, and pricing page.