Memtether

What memtether does

The memtether MCP server provides a local memory hub for multiple AI clients. Instead of maintaining separate memory stores, connected clients use one SQLite database, with file-level pointers linking each client to the shared file. The deployment does not require a cloud service, API fees, or a synchronization daemon.

Memory records retain their history. When a fact changes, the newer value supersedes the earlier one rather than deleting it. Records include who wrote or changed them, when the information was valid, and when the system recorded it. Conflict candidates can be reviewed and resolved by a human, with the decision retained in the record.

The project is listed in the 🧠 Knowledge & Memory category and is aimed at developers who need inspectable, local agent memory rather than an opaque overwrite-based store.

How it works

MemTether stores memory and audit data in SQLite. Its bi-temporal model separates the period when a fact was true from the time the system learned or recorded it. Supersession chains connect corrections to the facts they replace, allowing a developer or agent to inspect the full timeline.

Audit entries are linked through a hash chain. Periodic anchors cover batches of audit entries; verification reports whether an anchored section still matches its recorded hashes. Direct changes to covered database entries can therefore be detected. The repository describes this as a record-keeping design related to EU AI Act Article 12, not as certified compliance.

Retrieval includes deterministic SQL routing through SGM and hybrid embedding support. A standard installation can use keyword and BM25 search. Optional local embedding dependencies and model weights enable semantic search without sending data to a cloud service.

Setup and configuration

Install the package from PyPI and initialize it:

pip install memtether -i https://pypi.org/simple/
memtether init

Initialization creates the database, detects supported AI clients in standard installation paths, writes MCP configuration, and verifies the connections. Individual clients can instead be configured with commands such as memtether setup cursor or memtether setup claude-code. Custom client locations may not be detected automatically; memtether connect verify can confirm the result, and manual configuration is documented in the repository.

For semantic search, install the vector extra, download the models, and rebuild the index. The documented default model is bge-m3-int8, with a smaller Chinese profile also available. Omitting these steps leaves keyword and BM25 retrieval available.

The optional server extra adds the REST API and dashboard. memtether dashboard starts the dashboard. The REST API is unauthenticated unless MEMTETHER_API_KEY is set, so port 8080 should only be exposed to trusted networks.

Tools and capabilities

The memtether MCP server supports the memory workflows represented by the project’s CLI and examples:

• Create memories with remember, including a source identity.
• Search stored facts using keyword, BM25, deterministic SQL routing, or optional local embeddings.
• Correct a fact with correct, preserving the earlier value in a supersession chain.
• Inspect changes over time with timeline.
• Review a fact’s provenance, conflicts, temporal fields, and integrity status with court.
• Detect and record conflict decisions through the conflict workflow.
• Verify the tamper-evident audit chain.
• Migrate mem0 JSON exports through the documented Exchange v1 example without requiring the mem0 SDK.

Limitations and notes

Automatic client detection is documented for standard installation paths and may miss clients installed elsewhere. The REST API has no authentication by default. Semantic retrieval requires additional packages, model downloads, and index construction; without them, search falls back to keyword and BM25 methods.

Hash-chain coverage is batch-based, so recent writes can remain in an unanchored tail until the next anchor is created. The project’s performance figures are local SQLite microbenchmark results rather than a guarantee for every workload. Its Article 12 discussion describes design intent, not legal or regulatory certification.