Shared memory for multi-agent systems

The persistent memory layer
between your AI agents

Recall gives every agent one durable, shared memory. Write once, read anywhere. No more re-explaining context across handoffs, sessions, or models.

Open API · MCP native · Self-host or cloud

recall · shared memory graph
Agent A
Researcher
writes findings & sources
Agent B
Planner
reads context, sets tasks
recall core
Memory Layer
semantic + episodic store, shared across every agent
Agent C
Writer
recalls plan & facts
Agent D
Reviewer
audits full history
Drop into any stack MCP REST API Python SDK TypeScript SDK Webhooks Self-hosted
How it works

One memory, shared by every agent in the loop

Recall sits between your agents and their models. Every interaction is captured, structured, and made retrievable, so context survives across agents, tools, and time.

01 / Capture

Write once

Any agent commits facts, decisions, and artifacts to the shared store with a single call. No schema wrangling required.

02 / Structure

Auto-organize

Recall embeds, links, and deduplicates memories into a graph, blending semantic search with episodic timeline recall.

03 / Recall

Read anywhere

The next agent, session, or model queries the same memory and picks up exactly where the last one left off.

Built for developers

Three lines to give an agent a memory

A single endpoint for writes and reads. Recall handles embeddings, ranking, and retention so your agents stay focused on the task, not the plumbing.

  • Namespaces to scope memory per user, team, or run
  • Hybrid semantic and keyword retrieval out of the box
  • Automatic decay and pinning to keep context relevant
# give any agent shared memory
from recall import Memory

mem = Memory(namespace="team-atlas")

# agent A writes
mem.add("User prefers dark mode")

# agent B reads, later
ctx = mem.search("ui preferences")
print(ctx.text)
# -> "User prefers dark mode"
Memory graph

Every fact, connected and inspectable

Recall stores memories as a living graph you can browse, filter, and audit. See what each agent knows, where a belief came from, and how context flows between them.

  • Full provenance on every memory, down to the source call
  • Time-travel to any point in a run's history
  • Redact or correct memories without breaking references
Capabilities

Production memory, not a prototype

Everything you need to run shared memory for real agent workloads.

Isolation by default

Namespaced, encrypted memory with per-tenant boundaries so agents only see what they should.

Sub-50ms recall

Hybrid vector and keyword retrieval tuned for the latency budget of an agent loop.

Model agnostic

Works with any provider or framework. Swap models without losing a single memory.

Automatic retention

Decay, dedup, and pinning keep the store lean so recall quality never degrades.

Self-host or cloud

Run Recall on your own infra or use the managed cloud. Same API, your choice of boundary.

Observability

Trace every read and write, measure recall hit rates, and debug agent context live.

By the numbers

Memory that keeps up with your agents

<50ms
median recall
1
shared store per team
agents per namespace
99.9%
uptime target
Start free

Give your agents a memory
that lasts

Spin up a shared memory layer in minutes. Free to start, no card required.