Portable AI memory as MCP tools: five reads and three controlled writes.
MemoWeft MCP Server (io.github.memoweft/memoweft)
This MCP server exposes Portable AI memory as MCP tools. It provides five read operations and three controlled write operations to persist and retrieve long-term memory for TypeScript AI applications. Memory records can reflect what users said, what systems observed, what models inferred, and what remains conflicted.
🛠️ Key Features
Five reads
Three controlled writes
Separate records for user statements, system observations, model inferences, and conflicts
Portable long-term memory for AI applications
🚀 Use Cases
Agent memory for AI agents that need long-term context
Local-first semantic memory backed by a persistent store
User-profile memory in Node.js/TypeScript workflows
⚡ Developer Benefits
Tooling for integrating memory-store capabilities in MCP
Open-source long-term memory engine targeting TypeScript AI applications
Supports a SQLite-based approach (as indicated by topics)
⚠️ Limitations
Server description only specifies read/write counts; no other behavioral guarantees are provided in the source excerpt.
Give your AI a memory—without turning its guesses into your facts.
MemoWeft is an open-source long-term memory engine for TypeScript AI applications. It keeps what users said, what systems observed, what models inferred, and what remains conflicted as distinct records—so memory can be inspected, corrected, managed, and moved between hosts in SQLite controlled by your application.
MemoWeft is a library your application imports—not a chat product, hosted memory service, persona framework, vector database, or agent framework.
Why AI memory needs a paper trail
AI can already hold a convincing conversation. What it often lacks is reliable continuity.
Across conversations, important context can disappear. New information may quietly replace old information. A model's guess may return later as if the user had stated it. Move to another model or host, and the accumulated memory may be left behind.
MemoWeft does not ask a model to declare the truth. It preserves where information came from, when it appeared, what contradicts it, and why a memory was formed—so applications and users can inspect the path from evidence to recall.
Evidence stays evidence
User statements, observations, tool results, and model inferences keep distinct provenance.
Conflict stays visible
Corrections retain history. Unresolved contradictions are exposed instead of silently overwritten.
Memory stays yours
Hosts can inspect, manage, export, validate, and import versioned memory bundles.
Confidence is computed by rule rather than copied from a model's self-assessment. Transient states can age faster than durable facts and preferences. Built-in ingestion paths do not turn an assistant's own reply into user evidence simply because the assistant said it.
git clone https://github.com/memoweft/memoweft.git
cd memoweft
npm ci
npm run build
node examples/no-key-demo.ts
After dependencies are installed, this deterministic demo needs no API key, makes no network calls, uses an in-memory database, and writes nothing to disk.
text
[limited ] conf 600/1000 The user lives in Osaka — stated memory
[conflicted] conf 480/1000 The user lives in Tokyo — conflict kept, not overwritten
[candidate ] conf 200/1000 The user probably works somewhere central — guess (low confidence)
Summary: 3 cognitions, 1 in conflict-exposed state; inference remains labeled and rule-scored separately from stated memory.
Done. (in-memory database — nothing written to disk)
This proves MemoWeft's memory rules; it is not a model-quality benchmark. For correction history and typed decay as well, run npm run demo.
WeftMate is a desktop product built on MemoWeft. It turns the Core memory model into a visible profile, source trail, conflict view, and user-facing controls.
The screenshots below show WeftMate's UI, not UI bundled with MemoWeft Core. MemoWeft provides the memory layer and portable data contract; product experience remains the host application's responsibility.
Know what a memory came from. Each item can expose its type, confidence tier, and source trail.
Resolve short replies before storing them. A brief confirmation does not promote the assistant's proposal into user evidence.
On Node 20 or 22, also install the optional SQLite driver:
bash
npm install better-sqlite3
Save as quickstart.mjs:
js
import { createMemoWeftCore } from'memoweft';
const core = createMemoWeftCore({ dbPath: ':memory:' });
await core.ingestUserMessage({
subjectId: 'alice',
content: 'I only drink decaf after 3pm—caffeine wrecks my sleep.',
});
for (const item of core.memory.listEvidence({ subjectId: 'alice' })) {
console.log(item.sourceKind, '·', item.rawContent);
}
core.close();
Run it:
bash
node quickstart.mjs
Storing and reading raw evidence needs no model or network. Turning evidence into a profile, separating guesses from stated facts, and recalling it into later conversations requires a chat model. Embeddings are optional; without them, Core normally uses local FTS5 keyword recall.
MemoWeft keeps the journey from source material to recalled context explicit:
text
user words · observations · tool results
│
▼
evidence
│ provenance retained
▼
event
│
▼
cognition ◀── corrections and conflicts
│
▼
recall
The supported application path is the createMemoWeftCore() facade. Lower-level exports exist for advanced composition and carry documented stable, experimental, or internal support tiers.
Published packages and repository source move on independent release schedules. Check the installed release's npm metadata and package README for its exact compatibility range. Source previews are not presented as npm-installable until released.
MemoWeft is local-first through inspectable boundaries—not through a promise that data can never leave the device.
Memory is stored in an application-selected SQLite database; no managed memory service is required.
Raw evidence can be stored and read fully offline, and the repository includes a deterministic no-key demo.
Profile formation needs a chat model. Hosts may use a cloud or local OpenAI-compatible endpoint.
allowCloudRead filters evidence for MemoWeft's built-in cloud write-model prompts. It is not access control and does not govern custom code, recall, MCP tools, adapters, exports, or logs.
Observations and tool results default to ineligible for built-in cloud write prompts, but the host still owns consent, review, and authorization-change flows.
MemoWeft does not encrypt the SQLite file. Authentication, tenant isolation, encryption at rest, backups, logging, and compliance remain host responsibilities.
Forced removal of one evidence item removes its dependent derived events and cognitions and leaves an audit tombstone; it is not per-row physical erasure. Use resetSubject for a subject-level clear, and handle external indexes, logs, and backups at the host layer.
CI verifies offline regressions, API snapshots, runnable documentation snippets, builds, and Node compatibility. Published evaluation results document both their methodology and what they do not measure.
MemoWeft is library-first, and Core 1.0 is the first stable release of its supported TypeScript facade and memory contract. A plain npm install memoweft follows the stable latest line.
Stable, experimental, and internal surfaces are documented separately. After 1.0, breaking a stable symbol requires a major release and prior deprecation; experimental interfaces may still change in a minor release with notice. The Python package remains an experimental parity implementation rather than a feature-complete stable SDK.
Now: maintain the Core 1.x contract, expand versioned integrations, preserve Node 20/22/24 coverage, grow reproducible evaluation artifacts, and complete portable-bundle parity across TypeScript and Python.
Issues — reproducible bugs and concrete feature requests
Contributing — development setup and review expectations
Support — where to ask and what information to include
Contributions are welcome beyond Core code: clearer examples, framework integrations, platform testing, reproducible evaluation cases, and reviews of provenance, conflict, deletion, and privacy boundaries.
If you believe AI memory should be traceable, correctable, and portable—not an invisible black box—star MemoWeft, run the offline demo, or tell us what kind of memory experience you are building.
