Agent memory
Tune memory to the task, agent by agent
A recipe is a set of lever settings, so an agent can switch recipes when you ask in plain words, or autonomously on-the-fly, per task. A swarm does not have to run one compromise configuration: each agent fits its memory to the work in front of it.
At a glance
Recipes and levers are part of SHM, which comes with every paid plan.
SAIHM is built for enterprise-class agent teams, where each agent fits its memory to the work in front of it.
A recipe is a named set of lever settings.
Agents switch recipes when you ask in plain words, or autonomously on-the-fly, per task.
Recipes and levers are part of SHM, which comes with every paid plan.
Recipes and levers are part of SHM, which comes with every paid plan. Install it beside your agent:
npm install @saihm/shmSHM brings recall by meaning, recipes and levers, workstream continuity and compliance tooling, and it runs on your side. Gains vary by workload.
SHM · every paid plan
SAIHM is built for enterprise-class agent teams, where each agent fits its memory to the work in front of it.
SAIHM is built for enterprise-class agent teams, where each agent fits its memory to the work in front of it.
| Recipe | What it is for |
|---|---|
| Fast and lean | The quickest answers: a 26 ms median on our evaluation corpus. |
| Balanced | Recommended. Our best-measured ranking quality, about a quarter of a second at the median. |
| Deep recall | Reaches further into long or cross-referenced material, for research, audits and investigations. |
| Verbatim archives | Sharper precision when memory holds transcripts word for word. |
| Many workstreams | Each line of work stays whole and separate, however many run in parallel. |
| Compliance first | Built so an auditor’s questions can be answered from the memory itself. |
| Fleet | Visibility and consistent behavior across many agents and models. |
Around the recipes, SHM keeps work from blurring and puts compliance in the memory itself. Every workstream keeps its own continuous memory, facts from one client never bleed into another, and a correction made once stays corrected. Erasure cascades into every summary derived from a memory, and you can ask what was known when a decision was made.
A recipe is a named set of lever settings.
A recipe is a named set of lever settings. This is the lever panel for the Deep recall recipe, as it appears on the home page:
| Lever | In Deep recall |
|---|---|
| Reranker reads full text | On |
| Query expansion | On |
| Pointer dereference | On |
| Thread-aware grouping | Off |
| Injection budget | Set in tokens |
The first four levers are switches; the injection budget is set in tokens.
Agents switch recipes when you ask in plain words, or autonomously on-the-fly, per task.
Tell your agent what the work needs:
“For this audit, switch your memory to the Deep recall recipe: have the reranker read full text, and turn on query expansion and pointer dereference. Go back to Balanced when the audit is finished.”
Agents switch recipes when you ask in plain words, or autonomously on-the-fly, per task.
Triage agents run Fast and lean, investigators Deep recall, the records agent Compliance first.
Triage agents run Fast and lean, investigators Deep recall, the records agent Compliance first. Tuning is per agent and per task, so one swarm runs all three at once:
- triage-agent
- Fast and lean
- investigator
- Deep recall
- records-agent
- Compliance first
The Fleet recipe gives visibility and consistent behavior across many agents and models. The swarms and fleets page covers signed store policies and scopes.
Add this standing instruction to your agent’s instructions, such as a CLAUDE.md file.
Add this standing instruction to your agent’s instructions, such as a CLAUDE.md file. It covers recipe and on-the-fly lever adjustments, and parallelism as appropriate for main agents, sub-agents and the sub-agents they spawn:
Invariant: Always liberally use SAIHM protocol and SHM to autonomously maximize aggregate session token economy and memory performance, including recipe or on-the-fly lever adjustments and parallelism as appropriate for main, sub-agents, and sub-agents spawned by sub-agents. Strictly avoid unnecessary prose.
Recipes and levers need SHM, included with every paid plan. Gains vary by workload; the open benchmark states its scope.
- Do I need a paid plan?
- Yes, for recipes and levers. They are part of SHM, which comes with every paid plan. Install it with
npm install @saihm/shm. - Which recipe should I start with?
- Balanced. It is the recommended recipe, with our best-measured ranking quality and about a quarter of a second at the median.
- Can an agent change recipes by itself?
- Yes. Agents switch recipes when you ask in plain words, or autonomously on-the-fly, per task.
- Do the speed figures apply to my data?
- The 26 ms figure for Fast and lean is a median on our evaluation corpus, and the Balanced figure is a median too. Your workload will differ, and gains vary by workload.
No card and no wallet; you need Node.js 20.19 or later.
No card and no wallet; you need Node.js 20.19 or later. Any MCP host that runs local servers can add SAIHM with one command, and SHM comes with every paid plan:
npx -y @saihm/mcp-server-pro free-joinKeep reading: swarms and fleets, agent sharing, enterprise-class memory and the quickstart.