Open-source agent infrastructure

Local AI memory
for your agents.

Store facts, preferences and task state on your machine. Give your agents memory that lasts across sessions.

Run it locally.

Build from source with Rust 1.98+.

Install

cargo install \
  --git https://github.com/maskjelly/instantKV \
  --locked instantkv

Start a memory service

instantkv start --dir my-local-memory

Runs on: Your machine
Default address: 127.0.0.1:8080
Local data: .instantkv/data
License: MIT

Save once. Find it next session.

Your app chooses what to save and what to add to context. instantKV keeps the records and indexes consistent.

Save a preference

instantkv remember "Prefer Rust for local tools" --key preferences/language \
  --topic preferences

Retrieve it later

instantkv recall --topic preferences

In another terminal, run cd my-local-memory first. Credentials load automatically.

What you keep

preferences/language

Prefer Rust for local tools

Topic: preferences
Also supported: Tags, event time, custom JSON metadata
Storage: Record and indexes in one redb transaction

Read the memory contract →

For changing facts, use the optional memory lifecycle controller. It adds source quotes, revision-safe corrections, conflicts, live profiles and forgetting with replay protection.

Use the interface that fits.

One storage engine. Run it as a local service or embed the Rust core.

MCP

Connect an agent

Five memory tools, plus KV and checkpoint tools. Native stdio or an HTTP-backed bridge.

CLI

Work from your terminal

Remember, recall, search, browse and forget. Inspect health and restore task checkpoints.

HTTP

Use a scoped API

Namespace permissions, revision checks, TTL and bounded requests.

Rust

Embed the core

Use the engine in process. Keep storage and retrieval beside your runtime.

Know what you are installing.

Local by default

Private credentials. Loopback binding. Configurable cache, quotas and query budgets. Storage and retrieval need no cloud API or embedding model.

Configuration →

Explicit retrieval

search uses bounded BM25 with English stemming. Literal filters use topics, tags and time. Inspect truncated and query_reduced.

Query limits →

Your runtime stays in control

Your runtime chooses facts and context. The optional controller handles their lifecycle. Native phone bindings and device benchmarks remain planned.

Current scope and roadmap →

Operate your own node

Keep the data directory across upgrades. Back up the database while the server is stopped. Logical quotas do not cap total disk or process memory.

Backups and upgrades →

Evidence, with its conditions.

Published runs measure retrieval, model answers and storage separately. Small score gaps do not establish a significant win.

95.13%Source-session Recall@10

All 500 LongMemEval-S questions. Retrieval score, separate from answer accuracy.

14.34 MiBLargest sampled server RSS

LongMemEval-S on Apple M4 Pro. Sampled process memory; peak RAM was not measured.

8.31 MiBRecorded native binary

macOS ARM64 release build. Physical ARM-board and phone tests remain pending.

Benchmarks and raw data →Read the engineering notes →

Make it part of your runtime.

Start with the local profile. Save a record, connect your agent, and check the query limits.

Start locally →Build with us ↗Read the hosting guide →