GitHub project

Memory Hivefor agent teams.

A small file-based memory layer for AI agents.Orchestrate first, optimize every prompt, recall only what you need.

Workflow 0 orchestration Prompt optimize HyperRecall bundles Skill auto-route

create this skill... 5 recipes for tacos created 4k logo for Tendril remember: no bland UI source notes saved deployment lesson learned

Simple onboarding

Start with three commands.

Install the hive, give an agent a memory silo, then run the health check. No dashboard ritual required.

Install

Creates the hive, first agent silo, and boot instructions.

curl -fsSL https://hive.neural-forge.io/
install.sh | sh

Add agents

Gives each worker a private memory space without changing the workflow.

sh ~/.memory-hive/memory-hive
add knox

Verify

Checks the install so agents can boot, remember, and submit lessons.

sh ~/.memory-hive/memory-hive
doctor

v2.0.0 Orchestrate

Orchestrate first. Ship less context.

v2.0.0 runs Workflow 0 on every substantive prompt: platform detect → prompt-optimize → HyperRecall bundle → skills ensure → orchestrate. Prefer Grok/Cursor workers; the IDE model stays the planner.

99.28% token reduction vs naive full-boot. 138.9× efficiency vs naive (tokens into the agent turn). Improved vs HyperRecall-only: 969 ≤ 1,190 bundle tokens.

memory-hive platform detect --json
memory-hive prompt-optimize "your prompt"
memory-hive orchestrate "your prompt" --json
memory-hive bench suite --json

Measured 2026-07-09 Naive full-boot vs v2.0.0 optimize+bundle.

Naivefull boot corpus

139,614

v0.3.2HyperRecall bundle

1,190

v2.0.0optimize + bundle

1,005

99.28%token reduction vs naive

138.9×efficiency vs naive

v2.0.0release on GitHub

Dual baselines: naive = index+registry+knowledge+distilled+silo heads; v032 = HyperRecall bundle; v2 = prompt-optimize + budgeted bundle. Methodology in memory-hive bench suite --json.

Why it works

Memory without a platform tax.

No dashboard lock-in. No vector database requirement. No hidden daemon. Just auditable files, clear ownership, and a curation path from raw lessons to trusted knowledge.

Private by default

Each agent owns its own log and memory, so context does not collide.

Shared when useful

Durable lessons move into the hive only after they prove reusable.

Human-readable

Markdown and shell scripts make the system easy to inspect and repair.

Extensible

Swarm and Control add routing and safety without changing the core install.

Operating layer

One core product, clear extensions.

See the hierarchy

◈Core

Memory Hive

The required layer: private silos, shared hive, curator workflow.

✦Extension

Hive Swarm

Routes multi-agent work across available machines and capabilities.

✓Extension

Hive Control

Adds owner approval, safety checks, and recovery notes for risky actions.