Smartware

// why smartware

Most memory was built for chatbots. Agents need more.

Agents act on what they remember, often for many users at once, inside a fixed context window. Smartware is the next step in frontier research on how autonomous agents learn and remember, and it closes the three gaps that hold agents back.

Stale facts

Ask "where does Maya live now?" and most memory layers return her old address and her new one, unranked. An agent acting on the wrong one makes a real mistake.

SmartwareEvery fact knows when it was true. Old versions are marked superseded, so your agent gets the current answer, or the answer as of any date.

Leaky recall

Filtering permissions after ranking quietly drops good answers, and a summary built from a private message can surface to the wrong user.

SmartwareAccess is checked before anything is ranked. A summary is never more visible than its most private source. Zero leaks in our multi-tenant test.

Context bloat

Recall returns ten hits and hopes they fit. No token budget, no citations, no way to say "nothing relevant here".

SmartwareAsk with a token budget and get a context pack that fits it, with a citation and a reason for every item, ready to drop into the prompt.

// how it works

One call. Exactly the context your agent needs.

Your agent calls recall with a question, a point in time and a token budget, from any model or framework over MCP or HTTP. Smartware returns the current facts that fit, each with its source and the reason it made the cut.

POST /v1/acme/recall

{
  "query": "Where does Maya live now?",
  "as_of": "2026-06-30",
  "budget": { "tokens": 1500 },
  "include": { "history": true,
               "explain": true }
}

// the principal comes from your API key,
// never from the request body

from smartware import Smartware

mem = Smartware(tenant="acme", key="…")

pack = mem.recall(
    "Where does Maya live now?",
    as_of="2026-06-30",
    budget=1500,
)
prompt = pack.render()  # drop into context

{
  "mcpServers": {
    "smartware": {
      "command": "smartware",
      "args": ["mcp"]
    }
  }
}

// tools: recall, remember, forget, profile,
// history, feedback, reflect

currentfact · 12 May 2026 · slack #growth

Maya: I moved to Lisbon last month. (last month = April 2026)

supersededfact · valid until Apr 2026

Maya lives in Paris. → replaced by the Lisbon fact above

event12 May 2026 · same thread

Sam: Nice! Is the Lisbon office open yet? · reply kept with its question

keywordmeaningthreadtime

1,248 / 1,500 tokenscoverage 0.71 · 3 citations

// foundations

Four rules at the foundation.

Written into the protocol and tested on every change. They never bend, so your agents can build on them.

Three clocks

Every memory separates when it was learned, when it happened and when it was true. Your agent can ask what was true on any date, or what it knew at the time it acted.

observed

happened

valid

Permissions before ranking

Access is checked before anything is scored, on every channel. One Smartware can serve every user and tenant your agents work for, and a private DM can't leak through a digest.

DM · hr only#general · all→summary · hr only

Explicit supersession

When a fact changes, the old version is marked superseded and linked to the new one. Nothing is quietly overwritten, so you can always see what your agent believed and why.

Forgetting that sticks

Retract, redact or purge, with a dry run first. Forgetting cascades to every summary built from that memory and blocks it from coming back, which is what deletion requests actually need.

// self-improving

An autonomous research loop at its core.

Smartware improves itself with autonomous research loops. Every week its agents read the frontier, turn your feedback and requests into hypotheses, run the experiments and keep only what beats the last best. Every loop leaves it a little smarter, and you get each gain by upgrading, with nothing to tune.

Preview · example cycle data

Research

Frontier papers on agent learning and memory

Market

New releases and benchmarks across AI

Product feedback

What breaks, slows or confuses people using it

Community

Feature requests and recommendations from you

Livesignals this cycle: 110

Only better ships

A change merges only if it beats the current version on the benchmarks and every MUST in the spec still passes.

Agents never grade themselves

The keep-or-revert call is plain, deterministic code, not the agent that wrote the change.

Humans sign the spec

Agents improve the engine freely. Changes to the protocol itself need a maintainer's approval, so your integration never breaks.

Every change has a reason

Each release links the research, feedback or request behind it, with the benchmark run and the diff, in public.

// the record

+14.4 points of recall in 13 weeks. All agent-made.

Each square is an improvement that passed the gates and shipped. Recent ones glow; older ones settle into the structure. Drag back through the cycles to watch recall climb.

Cycle 13 ·

this cyclerecentsettledin flight

LoCoMo recall within 1,000 tokens, by cycle

// this cycle

What the agents are improving right now.

The live board for cycle 14. Anything that doesn't beat the current version goes back to Propose with the run attached.

// shape the next cycle

Tell the agents what to build next.

Sign in with GitHub and file a request, feedback or a paper worth reading. A triage agent merges duplicates, replies within the cycle and adds it to the next round of signals. Votes decide what gets proposed first.

Top requestsvotes

The loop is Loopy, an open protocol for autonomous research loops.

// benchmarks

Ahead where agents need it most.

Every system tested under identical conditions: same encoders, same token budgets, tuning on dev splits only, each test scored once. Smartware's agents rerun all of it every week.

Benchmark	smartwareSmartware	Next best	Lead	Why it matters to your agents
Permission test · multi-tenant	97.7 · 0 leaks	81.0	+16.7	Serve many users from one memory without leaking or losing answers
LoCoMo · 1,152 questions	85.6	79.7	+5.9	The right evidence from long, multi-session histories inside a tight budget
HotpotQA-50 · held out	93.3	92.0	+1.3	Multi-hop answers on documents never seen in tuning

Permission test · recall with access enforced

Smartware

97.7

Next best

81.0

LoCoMo · evidence inside the first 1,000 tokens

Smartware

85.6

Next best

79.7

Recall is measured as gold evidence inside a fixed token budget, so systems that return facts, turns or whole pages are compared fairly. The harness, configs and every per-query result are open source.

// user-owned · agent-led · open

User-owned. Agent-led. Open protocol.

Your memory belongs to you: host it yourself or anywhere, and export every record losslessly at any time. Agents lead its development in the open, and the protocol is public: fourteen verbs, one record format and a test for every MUST. Any engine that passes works with every client and agent.

recall

Budgeted, cited context for a question

remember

Idempotent writes, one or many

forget

Retract, redact or purge, with cascade

profile

Always-loaded rules and preferences

history

Every version of a fact

reflect

Summaries and observations with lineage

feedback

Learn standing rules from edits

changes

Ordered feed for sync and audit

Core

Records, remember, recall, budgets, idempotency, unknown parameters rejected.

Temporal

Three clocks, as_of and known_at, supersession links, history.

Governed

Access before ranking, provenance closure, redaction cascade, audit.

Adaptive

Reflect with lineage, rule proposals, lossless export and import.

MCP · stdio or Streamable HTTP

smartware_recall    readOnly · idempotent
smartware_remember  idempotent
smartware_forget    destructive · asks first
smartware://acme/mem/{id}

HTTP · bearer token → principal

POST /v1/{tenant}/recall
POST /v1/{tenant}/remember
GET  /v1/{tenant}/memories/{id}/history
GET  /v1/{tenant}/changes?cursor=

// who it's for

For anyone running agents.

Agent builders

Add one MCP server and your agents keep context between runs, models and frameworks. Swap the model, keep the memory, and pick up a better recall engine every week.

App developers

Give every user of your app an AI that remembers them, from one deployment, with each user's and tenant's data isolated and permissions enforced before anything is ranked.

Enterprise teams

One governed memory for every agent in the company. A summary can't reveal a private channel, every claim cites its source, and deletion requests cascade everywhere.

// show it off

Built on Smartware? Wear the mark.

Put a badge in your README, docs or footer. The cursor blinks and the mark fires now and then, so people notice. The conformance badge tells other developers exactly what your memory guarantees.

HTML embeds the badge inline, so it animates anywhere. The Markdown loads the badge from smartware.software, ready for any README. Keep the wording and colours as they are.

The first-ever self-improving memory protocol for agents and apps.

Most memory was built for chatbots. Agents need more.

Stale facts

Leaky recall

Context bloat

One call. Exactly the context your agent needs.

Four rules at the foundation.

Three clocks

Permissions before ranking

Explicit supersession

Forgetting that sticks

An autonomous research loop at its core.

Research

Market

Product feedback

Community

Only better ships

Agents never grade themselves

Humans sign the spec

Every change has a reason

+14.4 points of recall in 13 weeks. All agent-made.

What the agents are improving right now.

Tell the agents what to build next.

New request

Top requestsvotes

Ahead where agents need it most.

User-owned. Agent-led. Open protocol.

Core

Temporal

Governed

Adaptive

For anyone running agents.

Agent builders

App developers

Enterprise teams

Built on Smartware? Wear the mark.

Give your agents a memory that keeps getting better.