#1 on LongMemEval. #1 on LoCoMo. The new state of the art.

The memory layer
for your AI assistant.

PiePiper gives Claude, ChatGPT, and Gemini a persistent memory that follows you across conversations and across assistants. Local-first. Privacy by design. Top score on every public agent-memory benchmark.

No credit card. Runs on your machine. Cancel any time.

96.8%
LongMemEval (500Q)
#1 globally
91.0%
LoCoMo (1986Q)
#1 verifiable
93.9%
Combined average
Best in the field

Your AI forgets you every conversation.

Every time you start a new chat with Claude or ChatGPT, your assistant starts from zero. Your preferences, your context, your decisions — all gone. You re-explain, every single time.

PiePiper remembers.

PiePiper sits between you and your AI. It captures what matters, recalls it when relevant, and lets you forget what you don't want kept. Your memory belongs to you, not to OpenAI or Anthropic.

How it works

Three simple steps. Nothing leaves your machine.

PiePiper installs as a local service. It connects to your AI clients through the open MCP standard. Your conversations stay on your computer.

01

Capture

As you chat with Claude, ChatGPT, or Gemini, PiePiper observes the conversation through MCP and identifies durable facts, preferences, decisions, and context worth keeping.

02

Encode

Memory is stored locally with semantic embeddings, full-text indexes, temporal anchors, and reconsolidation. When you correct or update a memory, the original updates rather than just appending a new note.

03

Recall

When relevance matches your current question, PiePiper surfaces just enough context to your AI. You see exactly which memory is being used and where it came from, and can forget anything with one click.

YOUR MACHINE nothing leaves this box Claude Desktop · Code ChatGPT Desktop · Web Gemini Desktop · CLI Cursor Editor + more via MCP MCP PiePiper engine Capture Embed Recall Forget no API calls · no third parties SQLite + WAL Tiered local memory encrypted at rest · runs offline Hot recent · uncompressed Warm indexed · structured facts Cold consolidated · compressed
Verified results

Top scores on the two benchmarks that count.

Independent academic benchmarks that test real long-term conversational memory. We publish our raw numbers and the configs that produced them.

LongMemEval (ICLR 2025)
96.8%
500 questions · 80–120K token contexts
↑ #1 globally — beats OMEGA, Hindsight, Mastra
LoCoMo
91.0%
1,986 questions · 10 conversations · 300+ turns
↑ #1 verifiable — beats Hindsight, CORE, Memanto

Compared with the field

Independent academic benchmarks. Combined average is the simple mean of LongMemEval and LoCoMo. PiePiper is the only system in the top three on both — and the highest score on each, individually.

System
LongMemEval
LoCoMo
Combined
Status
PiePiper
96.8%
91.0%
93.9%
#1 on both
OMEGA
95.4%
Verifiable
Mastra OM (gpt-5-mini)
94.87%
Research
Hindsight (Vectorize)
91.4%
89.6%
90.5%
Verifiable
Backboard.io
90.1%
Press release
Memanto
89.8%
87.1%
88.45%
Paper
CORE (heysol)
88.2%
Verifiable
Zep
71.2%
75.1%
73.15%
Verifiable
Mem0
49.0%
<75%
Verifiable

Sources: Hindsight benchmarks, LongMemEval paper, LoCoMo project, OMEGA benchmarks, Mastra OM. Updated 2026-05-02.

Features

Memory that works the way memory should.

Cross-AI memory

One memory layer for Claude, ChatGPT, Gemini, Perplexity, and any MCP-compatible assistant. Switch tools without losing context.

Local-first by default

Your memory lives in a SQLite database on your machine. Nothing is uploaded unless you opt in to encrypted multi-device sync.

Memory dynamics

Reconsolidation, interference, and decay — the things real memory systems do. When you correct a fact, the memory updates rather than stacking up contradictions.

Provenance trails

Every memory has a source. Click any recalled fact to see the original conversation, the date, the AI that produced it.

One-click forget

Forget anything by topic, person, or time-range. PiePiper handles the cascade: related entries, derived summaries, embeddings — all gone.

Reconsolidating memory

Real memory updates when corrected, fades when unused, and reinforces when accessed. PiePiper is the only system that models memory dynamics rather than treating it as static append-only storage.

Pricing

Free for a month. €10/month after.

No credit card to start. Cancel anytime. Local data stays yours forever.

Free trial
€0/30 days

Full access to PiePiper Personal for 30 days. No credit card required.

  • All Personal features unlocked
  • Cross-AI memory (Claude, ChatGPT, Gemini)
  • Memory dynamics + provenance
  • Local-first, single device
  • Email support
Start free trial

No credit card. No auto-bill. Just install.

Most popular
Personal
€10/month

Everything in the free trial, plus encrypted multi-device sync and unlimited memory history.

  • Everything in Free trial
  • Encrypted cloud sync across devices
  • Unlimited memory history
  • Priority email support
  • Memory health dashboard
Start with €10/month

Cancel any time, keep your local data.

Business
Coming soon

Team memory pools, SSO, audit logs, custom retention policies, and dedicated support for engineering teams running multiple AIs.

  • Team memory pools (selective sharing)
  • SSO (SAML, OIDC)
  • Audit logs & compliance exports
  • Custom retention policies
  • Dedicated success manager
  • SOC 2 / HIPAA-ready deployment
Join the waitlist

Get early access when it launches.

FAQ

Questions worth asking.

Does my data leave my computer?

No. PiePiper stores everything in a SQLite database on your machine. The paid Personal tier offers optional encrypted sync between your own devices, but even then we never see plaintext. If you want zero cloud, the free tier works forever as local-only.

What happens to my memories if I cancel?

Your local memory database is yours. Cancelling stops the subscription and disables cloud sync, but your data remains on your machine. We also export to JSON anytime for free.

Which AI assistants does PiePiper work with today?

Anything with MCP support: Claude Desktop, Claude Code, Cursor, Cline, Windsurf, Zed. ChatGPT and Gemini are added as their MCP support rolls out. The bridge layer is one config file per client.

Why should I trust your benchmark numbers?

We publish the raw configs, the run commands, and the JSON results alongside every score. LongMemEval and LoCoMo are independent academic benchmarks with public datasets and publicly-available evaluation scripts — the same scripts every other system on the leaderboard uses. You can reproduce any of our numbers.

How does this compare to ChatGPT's built-in memory?

ChatGPT's memory works only inside ChatGPT, lives on OpenAI's servers, and ends when you cancel. PiePiper works across every assistant you use, stays on your machine, and is yours to keep. We're a layer above any single provider's memory feature.

Is PiePiper open source?

No. PiePiper is a commercial product. The architecture, the memory dynamics engine, the cross-AI integration, and the benchmark-leading retrieval pipeline are all proprietary. Open-source memory libraries exist; they don't score where we do. We charge €10/month because that's what it costs to keep building the best memory system in the field.

What about teams and companies?

The Business tier launches later this year with team memory pools, SSO, audit, and on-prem options. Join the waitlist if you want early access — pilot customers get input on the feature priorities.

Try PiePiper free for 30 days.

No credit card. Local-first. Works with the AI you already use.

Get started