Smartware
The first-ever self-improving memory protocol for agents and apps.
Memory that stays current, respects permissions and fits the context window. Built by agents, and better every week.
Website: smartware.software
Memory and context for AI agents and SaaS platforms, shipped as a kit: a protocol, a production engine, one server
binary, typed clients, adapters for the memory APIs projects already call, connectors that carry source permissions
into memory, and the tools to prove all of it on your own data.
Smartware 2.0 is a rebuild from scratch. The earlier code (v0.8.1 and before) lives on at
technodotventures/smartware-legacy.
What makes it different, in one line each:
- Admissibility before ranking. Recall over a store returns exactly what it would over the same store with every
record you cannot read removed: scores, order and text (M67). Hidden records cannot leak through statistics.
- Time is first class. Every record has when it was learned, when it happened and when it was valid;
as_of and
known_at answer "what was true then" and "what did we know then". Newer facts supersede older ones explicitly.
- Provenance closes access. A summary of records you cannot read is a record you cannot read.
- Budgeted, cited context packs, ranked by a declared, versioned ranker, with
why for every item.
- Proof travels with it.
smartware doctor checks any endpoint against the spec by MUST id; smartware eval re-measures
recall on your data with your reader.
Start here (the adoption ladder)
| Rung | Time | What you do | Level |
|---|
| 1. Plug in | 5 minutes | Add smartware serve to your MCP client; paste the memory policy | L1 |
| 2. Wire in | 1 hour | Use @technodotventures/smartware-client or smartware_client, or an adapter; map your auth to principals | L2 |
| 3. Govern | 1 day | Turn on connectors so source permissions flow into readers; multi-tenant; audit; smartware doctor --principals | L3 |
| 4. Adapt | 1 week | Reflect jobs, feedback, profile budgets, scheduled bundle exports | L4 |
Install a pinned version and check it (never a bare npx; nothing runs at install time):
npm install --global --ignore-scripts [email protected] # in a project: npm install --save-exact --ignore-scripts ...
npm view [email protected] dist.integrity # must equal the line for smartware-2.0.0.tgz in the release's INTEGRITY file
npm audit signatures # registry signatures and provenance: built by this repo's release workflow
Each release publishes INTEGRITY (the sha512 of every package tarball, as npm records it), SHA256SUMS of every
published file and a CycloneDX SBOM. The packages have no third-party runtime dependencies.
MCP (Claude, Cursor, Hermes, OpenClaw, any MCP client), local:
{ "mcpServers": { "memory": { "command": "smartware", "args": ["serve", "--store", "/Users/you/.smartware/smartware.db"],
"env": { "SMARTWARE_PRINCIPAL": "you", "SMARTWARE_TENANT": "personal" } } } }
A server for an app or a team (HTTP binding and MCP Streamable HTTP on one port):
smartware keys add --keys keys.json --id alice --tenant acme --groups sales # prints a token once; only its hash is kept
smartware serve --http 8787 --keys keys.json --store /var/lib/smartware/smartware.db # or --oidc-issuer https://idp... for OAuth
From code:
import { SmartwareClient } from "@technodotventures/smartware-client";
const memory = new SmartwareClient({ url: "http://127.0.0.1:8787", getToken: () => process.env.SMARTWARE_TOKEN! });
await memory.remember("Alice prefers oat-milk flat whites", { kind: "preference" }); // private to the writer by default
const pack = await memory.recall("What coffee does Alice like?", { budgetTokens: 1500 });
Already on another memory API? Keep your code: the MCP memory server tools (--tools mcp-memory),
a mem0-compatible REST API at /mem0, the mem0 Python class, a LangGraph
store (JS and Python), Vercel AI SDK middleware, Mastra tools and processors, and a Letta .af importer.
Moving in: smartware import-from mcp-memory|mem0|letta <file> --tenant <t> --as <user> reads the other project's file as
data. Moving out: smartware export --tenant <t> --out acme.smartware writes a lossless bundle any Smartware engine can import.
The kit
| Part | Where | What it is |
|---|
| Spec | spec/Smartware-0.1.md, spec/MUSTS.md | The protocol; every MUST has a stable id |
| Schema pack | schema/ | JSON Schemas, OpenAPI 3.1 and MCP tool schemas, generated from one source and drift-tested |
| Engine | packages/core | TypeScript product engine on Node's built-in SQLite. Zero runtime dependencies |
| Server | packages/server | smartware: serve, doctor, eval, keys, backup, export, import, import-from |
| Clients | packages/client, clients/python | Typed clients; identity is never an argument |
| Adapters | packages/adapters | tools, mastra, langgraph, ai-sdk; Python: LangGraph, mem0, Letta |
| Connectors | packages/connectors | SDK, Markdown vault, Slack, ACL replay test kit |
| Policy pack | policy/ |
How it is verified
- Parity with the oracle: 100% identical packs on LoCoMo test (1,152 questions, three ranker configurations) and on
LongMemEval-S; the oracle's own conformance suite passes against the product engine; one store file is shared by
both engines with byte-identical exports. results/PRODUCT_PARITY.md
- Benchmarks: the product engine reproduces the published results within 0.5 points. On LoCoMo test it is ahead of
the best of the 15 systems on all 7 headline metrics; on LongMemEval-S (day-precision times) 5 better, 7 level,
0 worse; in a pre-registered blind QA it was level with or ahead of all 16 systems with two different readers.
Baselines are LLM-free re-implementations of each project's recall path, not the products themselves; read
results/METHODOLOGY.md before quoting numbers. results/RESULTS.md
- Conformance:
smartware doctor --self reaches L4 (29 wire-level checks). It is engine-agnostic (the Python oracle also
reaches L4) and catches deliberate violations (global statistics, ignored parameters).
- Performance: results/PERFORMANCE.md.
- Security and supply chain: SECURITY.md. No lifecycle scripts, no third-party runtime dependencies,
release gate in CI and as a pre-push hook, tag-only staged publishing with provenance, approved by a maintainer with 2FA.
Develop
npm run build # TypeScript (needs typescript and @types/node: npm install --ignore-scripts)
npm test # 100+ TypeScript tests
npm run test:py # oracle suite, oracle suite against the TS engine, Python client
node tools/release-check.mjs # the release gate
git config core.hooksPath tools/hooks
Licence: Apache-2.0 (see LICENSE and NOTICE). Status against the implementation design doc:
docs/STATUS.md.
The website lives in docs/ (index.html, served by GitHub Pages); the guides beside it are plain Markdown.
Built by Techno Ventures.
