Token And Cost Savings

HivemindOS saves tokens by making agents search, reuse, and remember before they generate. Instead of pasting whole folders or chat histories into every prompt, agents pull only the context a task needs.

For recall accuracy, speed, and scale results, see Shared Brain Memory Benchmarks.

Where the savings come from

The best habit: keep lasting context in Hive Superbrain, reuse a reviewed skill for repeated work, and attach only the sources the current task needs.

Search the shared brain

hive-brain answer "query"
hive-brain recall "query" --scope full-vault --limit 8

Use hive-brain answer for a short grounded answer and hive-brain recall for a ranked list of hits with paths. Recall checks typed Agent Memory first and only searches the full vault when that isn’t enough, which is far cheaper than filling the prompt with broad context.

Reuse before you build

• karpathy-guidelines asks agents to read the relevant code first, make small changes, avoid speculative abstractions, define how to verify, and stop when the task is done. Most token spikes come from fixing avoidable bugs and untangling unrequested changes, not from the first attempt.
• hive-assimilate searches, in order: your pinned repos, paths, notes, or skills; the shared brain; the current workspace; your project folders; local indexes; and public GitHub. It also logs what was rejected, so the next agent doesn’t repeat dead ends.
• hive-capability-search answers “what can this hive already do for this task?” It returns the chosen components, alternatives, delivery channels, required credential names, side effects, and gaps. For example, “build this with parallel agents” maps to the /swarm-goal slash command, which creates a Work Board task that agents can pick up and act on.
• Hive Fusion: hive-skill-fusion saves a useful procedure as a shared skill, hive-workflow-fusion plans multi-step work with dry runs and approval gates, and hive-aeon-fusion turns recurring work into a background duty that reuses a saved plan.

Practical Rule

For expensive tasks, the intended order is:

1. Recall with hive-brain.
2. Map capabilities with hive-capability-search.
3. Reuse source with hive-assimilate.
4. Follow karpathy-guidelines while editing.
5. Turn repeated work into a skill or workflow with Hive Fusion.
6. Check runtime usage afterward.

That order is the cost-control loop: retrieve less, choose better, reuse more, edit smaller, and remember the result.

Evidence

In a live run on June 7, 2026, the same question (“Explain how HivemindOS shared brain memory works and which command a raw agent should use before relying on prior context.”) was sent to openai/gpt-4.1-mini through OpenRouter twice: once with a broad context pack and once with a targeted Hive recall pack. Both answers named hive-brain answer "<query>". Provider-reported usage:

A June 8, 2026 chatbot-build run used 87.5% fewer tokens with Hive context (42,571 versus 5,307 total), but it does not count as a saving: the Hive answer named files and tests that don’t exist in the repository, so it failed the outcome check. Token counts only qualify when the answer is also correct.

Run the benchmarks yourself

These commands run from a HivemindOS source checkout:

Live provider run

./scripts/hive-env-run -- pnpm benchmark:e2e-token-savings
./scripts/hive-env-run -- pnpm benchmark:e2e-token-savings -- --provider openrouter --model openai/gpt-4.1-mini
./scripts/hive-env-run -- pnpm benchmark:e2e-token-savings -- --scenario brain-recall --repeats 3

It needs a provider key in your shared hive env and supports --provider openai, --provider openrouter, and --provider usepod. Results are saved under .outputs/benchmarks/ with usage counts, timings, response ids, context hashes, and outcome grades, but never API keys or full prompts. If a broad context pack is cut to fit the cap, the result is marked context.truncated: true. A comparison is marked claim-ready only when both conditions meet the repeat threshold and outcomes are comparable.

Deterministic context budget

pnpm benchmark:context-savings
node scripts/benchmark-context-savings.mjs --input-price-per-million 2.50
node scripts/benchmark-context-savings.mjs --json

This estimates prompt size as max(characters / 4, lexical tokens) for broad loading versus targeted Hive recall, skills, and source paths. It is not a live E2E provider-billing benchmark: it isn’t an invoice and doesn’t judge answer quality. The old pnpm benchmark:token-savings name still works. Add --input-price-per-million with a current provider price for a dollar estimate.

Example run on the HivemindOS repository:

The exact percentage matters less than the repeatable gap between loading broadly and retrieving the smallest useful evidence.