The Plivel memory layer

One memory layerfor every agentyou ship

Write what users say, recall what matters, fade what doesn't. Plivel is the part of your agent that knows who it is talking to.

Start buildingStart building See pricingSee pricing

Hold to write

wrote “Prefers aisle seats” · mem_8f2k · 41 ms

Four things Plivel does between two messages.

Everything your agent needs to remember

Why Plivel

01 / 04

recall(user="dana_k", q=")

MemoryRelevance

Prefers aisle seats, never exit rows0.00

pinned

Flies TAP from Newark, Star Alliance Gold0.00

3 recalls

Asked for a vegetarian meal in May0.00

decaying

3 of 12,408 memories0 ms

The right memory, before the model starts talking.

Hybrid vector and graph search runs next to your agent in 14 regions. Voice agents get context inside the first breath, not after an awkward pause.

Pinned · penicillin allergyPreference · aisle seatOne-off · parked at SFO lot C

Memories fade unless they earn their keep.

Every memory carries a relevance score that decays over time and climbs each time it is recalled. Pin what must never fade, like an allergy, and let one-off chatter archive itself.

agent.py

written mem_8f2k · merged with 2 · 41 ms

Bring your agent. Keep your stack.

Python, TypeScript or plain HTTP. No new framework, no schema to design first. Write what the user said, recall what matters, and Plivel handles embedding, dedupe and merge.

TRACE tr_91ce · recall(dana_k)0ms

auth + tenant

1.2 ms

embed query

6.1 ms

vector search · 14 shards

14.8 ms

graph expand · 2 hops

7.3 ms

rerank · top 40 → 3

6.0 ms

compose context

2.6 ms

REGION FRA-2 · 14 SHARDSEXPORT → OTEL

See exactly why your agent remembered that.

Each recall ships a span-level trace: embed, search, graph expand, rerank. Export to OpenTelemetry, or debug in the Plivel console with the prompt it fed the model.

Life of a memory

One sentence, from the first word to cold storage.

Write

Your agent sends what the user said. One call, no schema.

Embed

Plivel turns it into a 1,536-dimension vector and strips anything marked private.

Link

The memory joins the user's graph, next to the facts it relates to.

Recall

Weeks later, a new question pulls it back in 38 ms.

Fade

Unused, its relevance decays. Every recall tops it back up.

Three lines in the language you already use

SDKs are open source and typed end to end. Everything else speaks plain HTTP or the Model Context Protocol.

PYTHON 3.9+

Python SDK

pip install plivel. Sync and async clients, Pydantic models for memories.

NODE · BUN · DENO

TypeScript SDK

Streaming recall, edge runtimes, typed metadata from your own schema.

GO 1.22+

Go SDK

Context-aware client with retries and connection pooling built in.

HTTP/2

REST and gRPC

Every endpoint documented with runnable curl and grpcurl examples.

MODEL CONTEXT PROTOCOL

MCP server

Give any MCP-capable agent memory tools with one config line.

OTLP

OpenTelemetry

Spans for every write and recall, GenAI semantic conventions.

SIGNED PAYLOADS

Webhooks

Get notified when memories merge, archive or are erased.

HCL

Terraform provider

Workspaces, keys, regions and decay policies as code.

Global by default

Memory that lives next to your agent

Each user's memory shard is placed in the region closest to where your agent runs, and pinned there if the law says it has to stay. Failover is automatic.

Regions

0 ms

p50 recall

0.00%

Uptime SLA

MEMORY SHARDS · 14 REGIONS

FRANKFURT · P50 RECALL 25 MSDATA RESIDENCY: EU · US · APAC PINNABLE

Security

Private by default, erasable on request

Trust center ↗Security docs

AES-256 AT REST · PER-TENANT KEYS · TLS 1.3

SOC 2 Type II, GDPR, HIPAA

Audited yearly. Reports and our BAA are one click away in the dashboard.

Residency you can pin

Keep EU users' memories in Frankfurt or Stockholm, and prove it with region-locked keys.

Tenant isolation

Every customer gets separate encryption keys and shards. No shared indexes, ever.

Erase with a receipt

One call deletes a user everywhere within 60 seconds and returns a signed receipt.

Benchmarks

Faster, cheaper, and it remembers the right thing

Measured September 2026 on 10M synthetic memories across 40k users, same embedding model for every setup, recall from fra-2. Scripts and raw numbers are in the docs so you can rerun them.

p95 recall latency at 10M memories · lower is better

Plivel

0 ms

Vector DB + Postgres glue

0 ms

Re-summarise history with an LLM

0 ms

Questions

Product questions

No. A vector store is one part. Plivel adds per-user isolation, a memory graph, decay and pinning, merge and dedupe, traces, and erase-by-user for GDPR. You would otherwise build those around the database yourself.

Any. Plivel returns plain text context and structured memories. It works with OpenAI, Anthropic, Mistral, Llama or your own fine-tune.

In the region you choose, encrypted at rest with per-tenant keys. EU-only and US-only residency are one setting. We never train on your data.

One call: DELETE /v1/users/:id. Memories, vectors and graph edges are erased within 60 seconds, and you get a signed receipt for your records.

Recall has a 99.99% SLA and fails over across regions. Agents can also run with a local read cache that serves the last known context.